------------------------------------------------------------------------------- ** Message sent August 1992 ** to everyone who requested PVM source code from netlib since February 1992. ------------------------------------------------------------------------------- >> Dear Colleague, >> In order to improve PVM we are trying to understand how >> people are using the system. >> Could you take a moment to answer three questions. >> >> Are you using PVM today? >> >> What are your application(s)? >> >> How many and what kind of machines have you used with PVM? >> >> Any comments and suggestions for improvement are welcome. >> Please send the information to pvm@msr.epm.ornl.gov. >> Your help is much appreciated. >> >> Best wishes, >> The PVM Project Team >> ------------------------------------------------------------------------------ From: swh@nx-01.lanl.gov (Stephen W. Hodson) 1.) yes, extensively 2.) numerically intensive computations on cluster of 16-IBM RS6000/560 connected with ethernet, fddi, and socc(IBM). PVM appears to be the software of choice because it is so simple to install and use. Steve Hodson Computing and Communications Division MS B-260, (505)665-4609 Los Alamos National Laboratory Los Alamos, New Mexico 87545 ------------------------------------------------------------------------------- From: wm@alvin.llnl.gov (Bill Meyer - x36039) > Are you using PVM today? Yes > > What are your application(s)? Computer aided tomography of the soft x-ray emission on the DIII-D tokamak. > How many and what kind of machines have you used with PVM? HP 9000 series 300, 400, and 700. Sun sparc. Solbourne (Sun). Cray. I've run up to 10 hosts at once. > > Any comments and suggestions for improvement are welcome. > Just got Hence. Looks good. Bill Meyer Lawrence Livremore Nat Lab ------------------------------------------------------------------------------- From: nrl@SSESCO.com Answers to questions: Yes we are using pvm right now (2.4.1). Our applications are large (100,000-500,000 line) scientific applications involving meteoorology, photochemistry, particle dispersion and underground transpost and aqeuous chemistry. We use it primarily on our eight processor rs6000 cluster, but have also included our Sparcs I & II from time to time. Someday I will get the Ardent Titan version running. We use pvm, Linda, p4 and express for a variety of comparisons, with pvm being preferred then Linda. I just completed 370 runs (requiring over 80 wall clock hours) of comparisons of pvm, Linda and express on our particle model. Run lengths ranged from five minutes to three hours wall clock time. I wrote an informal discourse on this bit of masochism for Dennis Duke at SCRI, because a paper he asked me to review, which purposrted to do a similar comparison, appeared to be badly flawed... it was. If you are interested I could expunge the disparing remarks about express and send you a copy in postscript format? -- Neil Lincoln | Mail: SSESCO Internet: nrl@SSESCO.com | Business And Technology Center Phone (612) 342-0003 | 511-Eleventh Avenue So. #216 FAX (612)-344-1716 | Minneapolis MN 55415-1536 ------------------------------------------------------------------------------- From: tchen@rayleigh.Mines.Colorado.EDU (Chen Tong) Thanks for the PVM programs. I used PVM last school year. Professor Dave Hale, who is now with Advanced Geophysical Inc., and I here at Colorado School of Mines tried some applications in seismic data processing. Our consortium group, the Center for Wave Phenomena, is aimed at trying new techniques for the oil exploration industry. Before we dicided to use PVM in seismic applications, we did some benchmarks comparing PVM and Linda, a commercial product. Our conclusion convinced us to choose PVM. One of our applications is running seismic migration on several workstations. Seismic migration, which gives the subsurface image of the earth from the data recorded at the surface, is computational intensive, especially in 3-D case. In real 3-D seismic migrtion, the data set is about 4G bytes and the computation required is A*1000**4 floating point operations, where A can reach about 30 or even around 100. Due to the limited disk space, we only tried the small size version of the problem, from which we concluded that we can run the real problem on 5 IBM/RS6000-520 and 530 in about 15 hours, which is comparable to the performance on CRAY. In this application, PVM is used not only to do the computation in parallel, but also to coordinate the disk I/O in order to make the data I/O in parallel. When testing PVM, we used up to 28 machines, including IBM/RS6000 and NeXT. In the seismic migration application, we used five IBM/RS6000 workstations. It seems to me that I get more problems using V2.4 than V2.3 maybe just I messed up on some machines. SUGGESTION: I got HeNCE after I coded up the application. The graphic display in HeNCE is very impressive. But my PVM version program can not do that. I wonder if we can have both the flexibility of PVM and the user interface of HeNCE. That will make PVM applications more attractive. Maybe Xab is the way, but I still have not make it running. Tong Chen ------------------------------------------------------------------------------- From: "Blaise M. Barney" > > Are you using PVM today? Yes - at the Cornell Theory Center, we've been using PVM for a good while now, and have also implemented PVMe on our RS6000 cluster SOCC switch. We have taught PVM during our Summer Program for Undergraduate Research (SPUR) and at least two students that I know of used PVM for their projects. Additionally, we've several staff using PVM for various reasons and several users also. We currently are planning a PVM workshop for October. > > What are your application(s)? That I know of: -Plasma physics -Volume rendering -Seismic Imaging -Linear Algebra algorithms -BLAS -Monte Carlo -Training and education - simple examples > > How many and what kind of machines have you used with PVM? Mostly IBM RS6000s in ethernet, token ring and SOCC networks. 7-8 machines have been available up until now, with plans to add more (8-16) in the near future. -Blaise Barney Scientific Computational Support Cornell Theory Center ------------------------------------------------------------------------------- From: usimap@sneffels.usi.utah.edu (Michael Pernice) Yes, I am using PVM, nearly all the time. I am developing a parallel solver for nonsymmetric partial differential equations, which I hope to eventually integrate into a large-scale combustion model. Currently I use PVM on a network of 4 IBM RS/6000 model 520 workstations, connected with a token ring network. These processors will soon be upgraded to model 560s, and the token ring network will be replaced by an FDDI network. I have written up some of the preliminary results in a tech report; I would be happy to send you a copy if you like. I have also recently participated in a procurement that will bring 8 model 560s connected with an FDDI network to our organization. (The project described above is part of a separate, IBM-funded project.) We plan to offer PVM as the basis for network parallel computing on this configuration. We will also join the two clusters across campus via a fiber-based backbone network. Hopefully we'll have some results by the end of the year. We also plan to offer a course in parallel numerical methods beginning in January. Students will use PVM on an Ethernet-based network of DecStations and will also access the nCUBE2 at SDSC (if our allocation is approved.) So far a major drawback with PVM is the lack of a network debugger. Today I spoke with someone who just returned from the Gordon conference, and he tells me that you are developing one. Also, I recently obtained the source for Xab from Adam Beguelin via anonymous ftp. I haven't had the chance to use it yet but it looks very promising. Finally, reducing the overhead of message initiation is desirable (though it's understandable if it can't be done because of conflicts with other design goals.) Keep up the great work! If NSF or DOE needs any evidence that they should continue to support your activities, I'd be quite happy to contribute. ----------------------------------------+-------------------------------------- Michael Pernice | Utah Supercomputing Institute | 85 Student Services Building Bell-net: (801) 581-7708 | University of Utah Internet: usimap@sneffels.usi.utah.edu | Salt Lake City, Utah 84112 ----------------------------------------+-------------------------------------- ------------------------------------------------------------------------------- From: Dick Wilmoth > Are you using PVM today? Yes, we have been using PVM continuously since Nov. 1991. > What are your application(s)? > We are running flow simulations for a variety of problems in hypersonic, rarefied flows using the Direct Simulation Monte Carlo method (a particle method). Some of these simulations have run more or less continuously for several months on multiple SparcStations for a single problem. > How many and what kind of machines have you used with PVM? > We are using PVM on a regular production basis on 8 SPARC-2's and have run some benchmarks using up to 32 Sparcs (a mixture of 1's, 1+'s, and 2's). However, network bandwidth wasn't really good enough for 32 machines. We have also been able to use the same codes running on up to 8 processors of a YMP and up to 4 processors of an SGI 4D/340. All of these codes use Intel NX-style message-passing calls which are translated into an equivalent set of PVM calls through a small library we have written. That way, we can use the same DSMC source code for the Intel iPSC/860, SPARCStations, YMP and SGI and simply link in the translation library for the latter three to use PVM. > Any comments and suggestions for improvement are welcome. > Please send the information to pvm@msr.epm.ornl.gov. > Your help is much appreciated. > We have been very happy with the overall performance even on local ethernet- connected SPARCStations. Have had some problems with reliability of messages using TCP sockets (node-to-node) but none to speak of using UDP sockets (node-to-daemon-to-daemon-to-node). Would like to see the ability added to re-use barrier names. I use barrier like 'gsync' and find the need to use it frequently. For long running programs, I can easily use up a lot of memory with strings of barrier names just to keep them unique. Would like to see a way to resolve the issue of having single hosts with multiple CPUs (like SGI, YMP). Right now, 'pstatus' seems to return the number of hosts rather than the number of CPUs. Maybe an optional entry in the hostfile? like 'cpus=4'. However, above all, just keep PVM simple and uncluttered like it is now. Dick Wilmoth Aerothermodynamics Branch Space Systems Division NASA Langley ------------------------------------------------------------------------------- From: Thomas M. Eidson 1. Yes, I am using PVM quite extensively and have taught a few others in out branch (Computational Science Branch, NASA Langley) to use it. 2. I use PVM to connect a host progrom running on my workstation with calculation codes running on the Cray's. Eventually I will use it to control codes on the Intel parallel computers also. The Cray usage is rather interesting. Using PVM along with rdist, rsh, etc. I have a demo where I run on the Cray in an interactive mode without ever logging into the Cray directly. In the demo, I move the files to the Cray, compile, execute, and delete the files remotely. I use this demo in an effort to show others how to reduce some of the "trash" load on supercomputers. 3. The primary use of PVM is for a multi-discipline design and optimization code that we are implementing in a multi-computer environment. In the early stages the various codes are running on several workstations. Larger computers will be incorporated later. I have written a "monitor" library which runs on top of PVM. This was done to add flexibility (incase we replace PVM), but also to have the user make single calls to monitor rather than multiple calls to PVM. For example, each discipline code is controled by a driver which (once execution is begun) sits in a wait state (at a rcv) at the top of a task loop. A master code calls the monitor to select a specific task and the monitor relays the desired task request to the discipline code. The "task request" requires several PVM calls to relay the necessary information and complete the hand-shaking. 4. Machines used: Sun, Decstations, SGI, Cray ------------------------------------------------------------------------------- From: anne@hermes.chpc.utexas.edu > Are you using PVM today? The last time I used PVM was yesterday. > > What are your application(s)? They are parallel versions of S. Altschul's program package BLAST, a set of programs that compare DNA or protein sequences against databases. > > How many and what kind of machines have you used with PVM? > Usually I use 2 IBM RS/6000 550. Sometimes I use Sun SPARC workstations. I have never tried to use more than 3 machines at a time. > Any comments and suggestions for improvement are welcome. I have had working sockets versions of my programs, before I started to use PVM. Therefore, it was very easy for me to use PVM. Regards, Anne Juelich Department of Applications Research and Development Center for High Performance Computing The University of Texas System Austin, Texas ------------------------------------------------------------------------------- From: root@calvin.jci.tju.edu (Super-User) In response to your query on pvm applications. The Jefferson Cancer Institute is part of the Jefferson Medical College, Thomas Jefferson University. The institute conducts basic and applied research in areas which include pharmacology, microbiology, molecular biology, molecular genetics, and structural biology. We have endeavoured to take maximum advantage of information technologies in all of these areas. Technologies in wide spread use include: relational and object oriented databases, high performance 3d graphics, analytical tools for molecular modeling, emergent computation and genetic algorithms. PVM supports applications in many of these areas which share two common threads: 1) algorithms which lend themselves to emergent approaches; and 2) applications which integrate data and services from multiple machines. With respect to emergent algorithms: As clearly indicated in PVM's accompanying references, the system is aimed at "regular" crowd computations which emergent and genetic algorithms may or may not fit squarely into. However, we have found that PVM can play a very effective role here as a primitive message communication component which is merged with IPC for intra- host communications. A typical role for PVM here is to act as the "boot and priority control" message channel which is the only guarenteed way of communicating with "all" processing agents in a particular network. PVM is responsible for booting all agents and also for control and reporting. With respect to data and services integration: PVM is used as a single message channel to integrate the services of a number of conventional applications into a complex. As we learn how to derive more knowledge for our existing databases of biological information, we move into complex applications. In this context, complex or composite applications, are simply circuits of conventional applications which cycle until minimum acceptance criteria are met. Acceptance criteria are representative of specific concepts in biology which relate directly to the experiment at hand. In summary, we wish to pose our questions in terms of relevant biological concepts and view results in those same terms. I hope that the above gives you some idea of our applications, what, why and how. With respect to machines: We have configured a distributed network of machines which include: Convex C240 Silicon Graphics: 4D/480, 310's, 35's Sun: Sparc II's NCD Xterm's Apple Macintosh's Of the Unix host capable machines this represents 13 units. Of the workstation terminal devices there are some 100 additional units. PVM is used on all of the Unix host capable machines. I would like to thank the PVM Project Team for implementing PVM. It is a most interesting and useful facility. By providing it to the community you have stimulated the growth of distributed computing and many of us benefit. Thank you once again. Jim Averback Director, Laboratory for Applied Computing Jefferson Cancer Institute Thomas Jefferson University ------------------------------------------------------------------------------- From: davidr@george.lbl.gov (David Robertson [ITG]) >Are you using PVM today? I was using it several months ago and plan to use it again as part of a research exhibit at Supercomputing '92. >What are your application(s)? I am using PVM as part of a program that renders very large medical data sets. >How many and what kinds of machines have you used with PVM? I have used PVM with up to 16 Sun workstations. I plan on using it with a CM5 as well for Supercomputing '92. Comments: Are you going to implement PVM with XDR record streams as well as with memory streams. Using TCP with memory streams seems to be a waste (i.e. limited packet size, starting and tearing down a connection each time, etc.) David Robertson dwrobertson@lbl.gov (510) 486-5778 ------------------------------------------------------------------------------- From: ecb@utrccm.res.utc.com (Bud Boman) PVM Project Team> Are you using PVM today? Yes. PVM Project Team> What are your application(s)? We have three: (1) Code EAGLE, a generalized three-dimensional, multi-block, Navier-Stokes algorithm, presently being used at UTRC to analyze fundamental fluid dynamic phenomena such as shock/boundary layer interaction and vortex generation devices. A parallel version of EAGLE has been implemented with PVM on multiple singleton processors of two Alliant FX/2800's and on a network of HP9000 workstations. Selected optimal cases dominated by floating point computations have shown 88% efficiency. (2) The key kernel of code FREEWAKE, a three-dimensional CFD algorithm which models the wake structure of helicopter rotor blades, currently being used by UTRC and Sikorsky to calculate wake influence in several aerodynamic/blade response codes. A parallel version of the kernel has been implemented with PVM on multiple singleton processors of two Alliant FX/2800's and on a network of HP9000 workstations. Selected optimal cases dominated by floating point computations have shown 99% efficiency. (3) WILDCAT, an unsteady Navier-Stokes code used to model the airflow through a jet engine. We are porting this code to run on a network of IBM rs6000's. PVM Project Team> How many and what kind of machines have you used PVM Project Team>with PVM? Many. So far we have loaded PVM on: a Convex C3240, a CM-200, 2 Alliant FX/2000's, 22 IBM RS6000's and 7 HP 700 series and HP 9000's. Most of our work so far has been on homogeneous systems/networks. That is, although all of our machines are networked we have not yet tried running any applications on say, the IBM workstations and the Convex together. PVM Project Team> Any comments and suggestions for improvement are PVM Project Team> welcome. - I've encountered at least one situation in which the best organizing paradigm for the parallelized code was master/slave. I was hindered by the default round-robin processor allocation scheme used by initiate. The problem was that in some situations a copy of the slave process would be started on the the master processor. The timesharing that resulted between master and slave processes slowed everything down. (Sometimes the master process died, due to the size of the code in question.) Would it be possible to designate one processor as the master which is then skipped by initiate()? (It would of course still be accessible via initiatem().) - Have you ported (or considered porting) PVM to any of the 386/486 based Unixes? - An integrated graphical interface for performance analysis would be great. If something could be tightly coupled with PVM, such that only a compile/link-time flag would produce performance graphics in real time. We've looked into HENCE, but it requires reworking the code with new library calls. I would like to see something completely integrated into PVM so that no extra calls would be required in the code. - There was some talk of providing virtual shared memory with PVM. We could use this. Network Computing Group United Technologies Research Center Silver Lane E. Hartford, CT. 06118 Bud Boman -- ecb@utrc.utc.com Angela Mounts -- abm@utrc.utc.com Chuck Rothauser -- chr@utrc.utc.com ------------------------------------------------------------------------------- From: kevin@psunuke.psu.edu I have just finshed developing a PVM interface to link the SCDAP/RELAP5 and CONTAIN computer codes. These two codes are used for transient analysis on nuclear reactor systems. The machines used for the development and implementation of this interface were IBM RS/6000s, one 520 and one 320H. If you have any questions or comments you can easily reach me since I am now working @ ORNL in the Engineering Technology Division. The best place to reach me is a97@ljows1.etd.ornl.gov Kevin Smith ------------------------------------------------------------------------------- From: jacobs@netcom.com (Allan Jacobs) Dear sirs: Our product (Forge90) translates from serial Fortran to parallel Fortran, inserting calls to a run time library. The run time library makes use of parallelizing system calls provided by vendors. One option available to our users is to make a PVM-compatible run time library. When using the PVM-compatible library, users call a master program that initiates copies of their (client) program. One of our goals is to make the parallel version of the code have the same external behavior as the serial version from which it is descended. One side effect of the PVM that makes the behavior of the parallel program different and less convenient is that data for units 5 and 6 (standard in and standard out) are sent to and from the master pvmd's terminal (as you document). Is there any work around that would allow us to attach to the tty of our master program? I understand that there is a new version of PVM in the works. Is there any documentation that we can get that defines changes to PVM's behavior, capabilities, and interface? Allan Jacobs Applied Parallel Research ------------------------------------------------------------------------------- From: htn@dipl.rdd.lmsc.lockheed.com (H.T. Nguyen) My name is H.T. Nguyen, I am currently with Lockheed Palo Alto Reasearch Lab., in Palo Alto, CA. > Are you using PVM today? Yes. > What are your application(s)? Combinatorial Optimization (The Assignment Problems). > How many and what kind of machines have you used with PVM? 2 to 16 SUN SPARCStation 2. H.T. Nguyen htn@dipl.rdd.lmsc.lockheed.com ------------------------------------------------------------------------------- From: hubler@galaxy.lerc.nasa.gov (Dale Hubler) > > Are you using PVM today? Yes, but not personally. I am the sys admin for the Aeropropulsion Analysis Office at Lewis Research and have installed PVM on the RS6000, Sun, & SGI platforms here. I know of two separate projects at the lab using PVM. These are active efforts. > > What are your application(s)? The first application is a CFD problem and the researcher has a version which has worked fine on heterogenous clusters, some including a Cray as one of the nodes. This was a Cray application that was converted to run on a cluster of RS6000s. Turnaround time was dramatically improved by using the cluster. We are currently planning implementation of a 32 node cluster and this will be one of the applications. A paper was written and will be presented at a Cray User's Group conference in Washington this September. The second application is in the early stages and is a coarse breakdown of an optimization problem within an aircraft design and mission analysis simulation. Instead of running optimization cases sequentially on one processor we are trying to split the cases up across a cluster (again, RS6000s) using PVM. > > How many and what kind of machines have you used with PVM? We are primarily using RS6000, but we have succesfully included Suns, SGIs, and Crays in the cluster. Most dramatic improvement comes from sending the most intensive part of the problem to the Cray and allowing other parts to run on the RS6000 model 550/560s. > > Any comments and suggestions for improvement are welcome. > Please send the information to pvm@msr.epm.ornl.gov. > Your help is much appreciated. We have had some problems with the PVM daemon on one or another system becoming "unresponsive". There is no alternative to my knowledge of detecting this and adjusting the cluster that I know of. But again, I am not the programmer and could just be ignorant. But it would be nice to have some increased debugging and recovery options. Also, some performance analysis tools would be helpful. But overall we think the entire clustering concept is very exciting and this area is sure to evolve. Thank you very much for the tools we have so far. -- Dale A. Hubler (216) 977-7014 hubler@lerc.nasa.gov ------------------------------------------------------------------------------- From: eugene@nas.nasa.gov (Eugene N. Miya) PVM today: It has been brought up, we are trying to figure out how to administer the thing so that our users (nearly 1K) don't all try to run at once. 1-2 users have had experience else where and are chomping at the bit to use workstations and our network resource. Applications: CFD, and graphics/distributed program control. Many and kinds of platforms: Over 100 WKs are available, most the entire range of SGIs. Also brought it up on the Crays (Y and 2), Convex, and we have Intel cubes ready as well. We have problems with control and accountability with PVM. Our users are quite devious about their use of cycles and they could easily swamp our networks if they used PVM extensively. We are trying to figure out how to handle this problem. ------------------------------------------------------------------------------ From: lee@SLCS.SLB.COM > Are you using PVM today? YES. > What are your application(s)? Finite Element. > How many and what kind of machines have you used with PVM? SUN Sparc and IBM RS6000. ------------------------------------------------------------------------------ From: astfalk@hydra.convex.com (Greg Astfalk) > Are you using PVM today? Yes. > What are your application(s)? At the moment it is a finite-element code (2D). In the near future I will be doing a much larger application that involves multi-scenario optimization. That will use an interior-point based LP code. > How many and what kind of machines have you used with PVM? Convex (two differnet types), Hewlett-Packard workstations (several models) and Sun workstations. So far I've used up to 8 machines concurrently. The multi-scenario effort will hopefully use several dozen machines via Convex's `world-wide' WAN. ------------------------------------------------------------------------------ From: cljanss@snll-arpagw.llnl.gov (janssen curtis l) >Are you using PVM today? Yes. >What are your application(s)? Computional chemistry--Self-consistent-field theory energies and derivatives of the energy wrt displacements of nuclear coordinates. >How many and what kind of machines have you used with PVM? 18 RS6000 550's (AIX 3.2, pvm 2.4.1) connected by FDDI. 1 SGI 8 processor (33 MHz IP7) machine (IRIX 4.0.1, pvm 2.4) 1 SGI 4 processor (25 MHz IP7) machine (IRIX 4.0.1, pvm 2.4) >Any comments and suggestions for improvement are welcome My application is communications intensive at times. I haven't completed the RS6000 cluster benchmarks, but the multiprocessor SGI timings (real time as opposed to CPU time) were disappointing. Even when I didn't use the network and only communicated between processes on the same machine, my node processes seemed to spend much time waiting for the pvmd process to catch up. A version of pvm with more direct support for multiprocessors would be very useful. Curt Janssen Center for Computational Eng. Org 8117 Sandia National Labs Livermore, CA 94551 (510)294-1509 ------------------------------------------------------------------------------- From: lhdsy1!grug21.nor.chevron.com!jnrhi@uunet.UU.NET (N. R. Hill) Yes, I am using PVM today. I use it almost every day. I use PVM to process and analyze seismic data. In particular, I'm using it to construct 3 dimensional images of the earth's subsurface. So far, I have been working on Dec5000's. In a couple of months, I'll be trying out my code on some of our other workstations ( HP,RS6000,SGI) and on a Cray. I hope to arrange trying out my code on an Intel or a CM5. My main comment is that I've been very pleased with PVM and have had no problems. The provided functions are convenient for me. I'm looking forward to trying out PVM on other machines. It will be great if the same code will perform on a heterogeneous workstation network and on a machine like a CM5. Are there any PVM bulletin boards or user groups? Thanks for the software. It's been a real resource to me. Sincerely, Ross Hill ------------------------------------------------------------------------------- From: otto@iliamna.cse.ogi.edu (Steve Otto) > >Are you using PVM today? > Yes, to study the practicality of using workstations as a parallel resource. >What are your application(s)? > Scientific programs -- a multigrid PDE solver, a cyclic jacobi eigensolver. We are studying having MetaMP (a programming system of ours) compile down to PVM. We are also interested in developing a subset of PVM that would support multi-threading on each processor and lightweight context-switching so as to allow overlap of computation and communication (communication masking). Is the PVM project doing this? >How many and what kind of machines have you used with PVM? > Sun 3's, Sparcs, and HP workstations. >Any comments and suggestions for improvement are welcome. >Please send the information to pvm@msr.epm.ornl.gov. >Your help is much appreciated. You're welcome. --Steve Otto, OGI ------------------------------------------------------------------------------- From: Ram Rajagopalan In reply to your questions ... Yes, I am using PVM today. Our applications so far has been in Fluid flow and Monte carlo simulations. One requires a good bit of communication and the other requires little between nodes. We run primarily on Silicon Graphics Machines, although there were some instances where RS6000 were also part of the network. I would like to make a few suggestions regarding documentation. Currently it is very terse, I would like to see more examples using all the possible calls one can make to the pvm library. Also it would be usefule if rcv can be modified so that it blocks until a specific message from a GIVEN NODE is arrived as opposed to just a message type. Other than this I think the product is just peachy .... Regards, Ram Rajagopalan (908)730-2556 Exxon Research & Engineering Co. Clinton Township, Rt 22 East Annandale, New Jersey 08801 ------------------------------------------------------------------------------- From: forrest@watson.ibm.com I am using pvm at present. The application is Mixed Integer Programming using Branch and Bound. If you have the latest copy of OR/MS today you will see a small announcement that IBM has a limited release of OSL (Optimization Subroutine Library) using Parasoft's Express. I am now using pvm because it seems more reliable (if a bit slower) and has no licensing problems. I have used 28 RS 6000's (some in Dallas and Almaden) with no problems other than irate users. I have found one limitation (I think). If there are several messages outstanding and I probe for one particular one then it can return not found, even though a rcv finds it. From the code it looks as if probe gives up rather quickly. I have coding which allows the user to delete or add slaves as long as they are on original pvm list. This is useful but could be more useful if I could dynamically delete or add them to pvm. John Forrest ------------------------------------------------------------------------------ From: roy@nersc.gov (Roy Troutman) Are you using PVM today? Yes What are your application(s)? Ray Tracer. Radiosity. How many and what kind of machines have you used with PVM? NeXT Sun IPC ------------------------------------------------------------------------------ From: "Jean-Pierre Prost" 1. Yes, I am using PVM currently. 2. The applications I am dealing with are mainly fluid dynamics FORTRAN codes. 3. I am running PVM on clusters of IBM RS/6000 (between 3 and 6 machines) and I have one application running on a cluster of 5 IBM RS/6000 and one Silicon Graphics workstation. No special comments about PVM. Just a pretty neat package, quite reliable. Hope the answers to your questions are satisfactory. Let me now ask you one question. What is the current status of the announced debugger for PVM-based distributed applications ? Is it still under development, or is there any running version currently available. Please let me know. Thanks in advance, Jean-Pierre. ------------------------------------------------------------------------------- From: johnson@vaxino.scri.fsu.edu > >Are you using PVM today? YES > >What are your application(s)? A weather model. Specifically run the Advanced Regional Prediction System (ARPS) developed at the Univ. of Okla. > >How many and what kind of machines have you used with PVM? We have run PVM with a cluster of 17 RS6000's ------------------------------------------------------------------------------- From: U-E65718-Bob Hanks (1) Are you using PVM today? Yes, doing evaluations and demonstrations for engineering workstation users. (2) What are your application(s)? So far, we are working with the provided examples. However, I did rework the fractal plotting routines to use the HOOPS graphics library. The new main routine plots the strips of results as soon as they arrive from the processing nodes and allows one to rubberband a new fractal region for display. Later this week we will meet with engineers working on electromagnetic signature problems. They are interested in the possibility of parallelizing large, dense matrix solution methods. There has also been interest from our Fluid Mechanics people in using PVM to distribute CFD problems in Turbomachinery across workstations. (3) How many and what kind of machines have you used with PVM? Half a dozen HP workstations (710/HPUX8.07, 720/HPUX8.05, 400T/HPUX8.05). The 400T is a CISC machine so I defined a new architecture with no problems. I have also used PVM on our Cray YMP/UNICOS 6.1.6. The electromagnetics people will want to use it with DEC/Ultrix. ------------------------------------------------------------------------------- From: bartol@helmholtz.sdsc.edu (Tom Bartol) Dear PVM Project Team, Thanks for providing such a great tool and for requesting input for its continued enhancement. We acquired PVM only very recently to evaluate its usefullness in our heterogenous computing environment here in the Computational Neurobiology Laboratory at the Salk Institute. We currently have several computer simulation projects running and in development aimed at illuminating our understanding of brain function from the molecular level to the network level. We are developing our simulators to run on the nCUBE2, IPSC, CM5, KSR1, Sun Sparc2, IBM RS6000, and MIPS R3000 computers. We are using Parasoft's EXPRESS software to handle interprocess communication on the parallel machines. However, a common problem among our simulators is the need for transportable computer graphics output. Thus we are hoping that PVM will give us the necessary tools to build a generic interface wereby our simulators can communicate with different graphics rendering engines. For example, we would like to be able to have an SGI box intercept the graphics commands sent to it by a live parallel process (say on the IPSC) and then render the results on the fly. We envision interfaces via GL calls as well as X11R5 PEX calls. For us, PVM provides a missing link between the various machines we use every day. If I can give further information, or if you have any ideas floating around which might help us, please don't hesitate to contact me again. Thanx, Tom Bartol ------------------------------------------------------------------------------- From: "w.s. young" Yes, I am using PVM today. I am developping algorithms for molecular dynamics calculations on parallel machines. To date this has been done on DecStation 5000's using anywhere from 2 to 8 hosts. Since these codes are from CM-5's or models for CM-5 codes it would be nice to see a port that runs on the CM-5. Also, in the way of addtions it would be nice to have some higher level operations such as global sums, min/max and scatter/gathers. While these are relatively easy to contruct from the communications primitives given, I would expect that with a more in depth understanding of PVM than I have it would be possible to write far more optimized versions of these routines. William S. Young ------------------------------------------------------------------------------- From: trey arthur |> |>Are you using PVM today? yes. |> |>What are your application(s)? currently i have tested pvm on three benchmarks: Embarrassingly Parallel (Monte Carlo) one global is sum is done 3D potential solver multigrid NAS benchmark Hessenburg Matrix reduction several global sums done 3D Elliptic solver planes of data are transmitted each iteration Unstructed grid generator each machine creates a patch of the grid |> |>How many and what kind of machines have you used with PVM? Cray, Convex, CM-2, Intel/i860, Sun3, Sun4 IBM RS/6000, SGI workstations |> |>Any comments and suggestions for improvement are welcome. |>Please send the information to pvm@msr.epm.ornl.gov. |>Your help is much appreciated. |> DEBUGGER!!!! DEBUGGER!!!! DEBUGGER!!!! global routines would be nice. i have written some on my own and if you would like, i can mail them to you. (the routines emulate the Intel's global routines on a i860) update the PVM user's guide. (the one i have is over a year old). PVM is very impressive, esp for a free package! I have been very surprised by the performance I get. PVM seems to perform very well for problem sizes of varing granularity. i remember someone telling me there was going to be an X interface call XAB. is it still being developed? when is the next release of HeNCE coming out? later, trey *************************************************************** * Trey Arthur CSC NASA Langley Research Center * * Aerospace Engineer MS 125, Hampton, VA 23665 * * trey@godzilla.larc.nasa.gov (804) 864-8374 * *************************************************************** ------------------------------------------------------------------------------- From: sonnenfe@superc.nosc.mil (Bruce S. Sonnenfeld) Yes, we are using PVM today. The applications we are using it on include molecular modeling and tactical applications. We presently have a suite of 4 machines (a MasPar, Sequent, Wavetracer, and nCube). They use various front end CPU's such as HP 720, Sun 410, and Dec 5000. If you need any more info, let me know. Regards, Bruce ------------------------------------------------------------------------------- From: Phil Neal Inc. Are you using PVM today? YES YES YES. I am trying to implement some neural net stuff using PVM. In particular, I want to implement a dotproduct function and a matrix multiplication function. What are your application(s)? In actuality , I am trying to speed up calculations for the neural net stuff, which is interfaced from Splus. Eventually , I want to have the pvm stuff in some sort of library that can be called/accessed directly from Splus. How many and what kind of machines have you used with PVM? I am running on 4 Sparc 2s and a SUN IPC all running the same rev of Sun Os 4.1.1. I tried running on our SUN 3's as well but the sun 3's always seemed to slow things down too much. OTHER MACHINES I tried running it on our IBM 3090 or 4381 or some such thing compiling as if it was an AIX r6000 machine but I don't think it worked. It would compile. But it would never get back to me when I tried to send work to it. It i is also one of the public use computers here at the U and response time is sporadic. I also tried a few of the sequent/dec public machines. It compiled ok. But I can't remember what the response was. HENCE I have tried using hence. But I don't trust it. It is a nice interface, but it is sort of self-contained and I don't have the time to parse the code myself. I tried to compile up the dotprod demo (for hence) but it wouldn't run. It sort of locked up somewhere. IMPROVEMENTS HINTS ETC. More demo programs. How about a library of math routines like dotprod,matinv, matmult, random number generators, etc. HOW ABOUT LAPACK for PVM ???? Also, what are the advantages of broadcasting the data, like a big array to all the slaves as opposed to just the piece they will work on ? I think you are doing a great job. I love PVM. Good luck, I will have more questions for you as time goes by. Phil Phil Neal, Systems Programmer, Statistics Department GN-22 University of Washington, Seattle, Wa. 98195 USA | 206-685-1627 ------------------------------------------------------------------------------- From: hart@nipmuc.fsl.noaa.gov (Leslie Hart) > > Are you using PVM today? Yes, we are just beginning testing it. > > What are your application(s)? Meteorological modelling. > > How many and what kind of machines have you used with PVM? Sun Sparc2. KSR/1. I just built PVM on the KSR/1 today. I was wondering if you had any experience there at ORNL on your KSR/1 with PVM. I just set the RSHCOMMAND to /usr/bin/rsh and it seemd to build and work OK. It also seems to work OK between our Sun Sparc2's and the KSR. The performance is a little slow on the KSR and I don't know why yet. ------------------------------------------------------------------------------- From: auerbach@xenopus.med.buffalo.edu (A. Auerbach) 1. Yes, we use pvm all the time. It is an integral part of our computing. 2. Our main pvm applications corcern kinetic analysis of ion channel currents. This problem is solved both by analytical and Monte Carlo simulation methods. 3. We use pvm on a network of Sun sparcs and IBM rs/6000. Our cluster of rs6000s now stands at 5 (4 model 560s and 1 model 970), and we expect to double or triple the number of nodes in hte next several months. We hope that IBM will announce a high-end multiprocessor system. Typically, we use 4 or 5 nodes for a job without loss of linear speedup. 4. PLease put us on your mailing list. We are very interested in updates, etc. WE find that the pvm parallel programming model is well sutied to our comutational tasks. Anthony Auerbach Dept. Biophysics. SUNY Buffalo 324 CAry Hall Buffalo, NY 14214 716-831-2435 auerbach@xenopus.med.buffalo.edu ------------------------------------------------------------------------------- From: jonathan@windchime.nersc.gov (Jonathan Brown) > > Are you using PVM today? > Yes. > What are your application(s)? > At NERSC, most of our PVM experience has been with small test programs, but we are starting to use it in a large atmospheric modelling code. Others here have used it for ray-tracing. > How many and what kind of machines have you used with PVM? > 32 Sun Sparc 18 IBM RS/6000 2 Crays 1 BBN Butterfly 1 HP 9000 1 NeXt > Any comments and suggestions for improvement are welcome. > Please send the information to pvm@msr.epm.ornl.gov. > Your help is much appreciated. > Our application previously ran only on the BBN using LMPS. With PVM we can distribute the computation among any of our machines, thanks to its availability and portability. One issue that came up when converting from LMPS to PVM is that the PVM rcv() call cannot ask for a message from a specific sender. We had to incorporate the instance number of each message in the message type argument. As we get more experience with PVM in our large application I will provide you with more feedback. ---------------------------------------------------------------------- Jonathan Brown jonathan@windchime.nersc.gov National Energy Research Supercomputer Center Lawrence Livermore National Laboratory P.O. Box 5509, L-560 Livermore, CA 94550 (415) 423-4157 ---------------------------------------------------------------------- ------------------------------------------------------------------------------- From: scroggs@wave.math.ncsu.edu (Jeff Scroggs) I forwarded your request on to Neil Sigmon--the real user of the system. In addition to Neil's current work, there are plans to test the use of pvm on a network of IBM's. Best, Jeff Scroggs ------------------------------------------------------------------------------- From: martin@wildhog.stanford.edu (Martin Rinard) I'm using PVM today. It is the message passing platform for one of the implementations of Jade, a language for writing serial, imperative programs that run in parallel on a hetergenous collection of machines. We have a sparse cholesky factorization algorithm, a program that simulates water, and an implementation of the barnes-hut n-body solver. I have used up to 6 machines (so far), these are PMAX, SGI, DASH (the Stanford DASH multiprocessor) and SUN4. Martin ------------------------------------------------------------------------------- From: dhinds@allegro.stanford.edu (David Hinds) > Are you using PVM today? Well, probably not today, but I do use it semi-regularly. > What are your application(s)? I'm developing a method for protein structure prediction, for which I exhaustively search all possible tertiary conformations using a simplified representation of a polypeptide chain. The code is essentially one huge tree search; it parallelizes easily, with the only overhead being some message passing to balance the load across different processors. I have a sort of parallel interface library that lets me run the same code on either a distributed-memory Ncube-2 system, a shared-memory Silicon Graphics 4D-240, or a network of workstations running PVM. > How many and what kind of machines have you used with PVM? Mostly a cluster of seven Silicon Graphics workstations, including Indigos, the 4D-240, and a Crimson. I haven't really tried mixing architectures. > Any comments and suggestions for improvement are welcome. Is there a way to indicate in the host file that multiple instances of a PVM program should be started on the same machine? This would be useful to me when I want to use PVM to spread a job over several workstations, but still want to take advantage of the four processors on our 4D-240. - David Hinds dhinds@allegro.stanford.edu ------------------------------------------------------------------------------- From: "Duane A. Bailey" > Are you using PVM today? Yes. Today. I use PVM as a replacement -- almost -- for the Cosmic Environment. > What are your application(s)? PVM is the target message passing system for a higher level communication system I've constructed. (PVM provides portability and support for heterogeneous environments, of course.) Applications on top of this include a large-prime detection algorithm, various combinatorial optimization codes, etc. I also use PVM for course support. > How many and what kind of machines have you used with PVM? Sequents (Symmetries (and a Balance [ha!])) and Networks of Sparcs. > Any comments and suggestions for improvement are welcome. 1. PVM is great. I use it all the time. 2. Make sure the header file is up to date. (e.g. pvm_perror) It should be installed. 3. Why are processes identified by strings? It seems it would induce a large overhead in a program that, for example, sends short messages. Provide an "int interface"! 4. Make sure that rcvinfo returns the number of bytes the *user* stuffed in the buffer, and the number of bytes the machines received from remote. Padding and byte-order manipulation seems to throw the number off. I've seen a number of users who were surprised by this. 5. The addressing scheme in PVM makes decentralized process initiation and communication difficult. Here's why: In order to send a message to a process, you must have its address, which consists of its "name" and its "instance". The programmer picks the name, and the system generates the instance. This means that to communicate with a process, you should have (1) created it, or (2) have been informed of the instance by the creating process, or (3) make an assumption about the value of the instance (say, it is 0). In the cosmic environment, addresses are generated as "processor"/"pid" pairs of integers. The user picks both, and informs the system what they are at spawning of the process. This has the advantage that for those algorithms that have considered the process structure ahead of time, they may address remotely created processes directly; they needn't have parts of the process address passed to them. A workaround, obviously, is to sprintf/sscanf a cosmic-like address into the "name" field, and assume that the instance numbers will all be 0 (which they should be). It seems like a lot of overhead, though. A suggestion: Suppose PVM had the following interface: int initiate(char *name, char *arch, int suggestedInstance) Suggested instance would be the instance number assigned (by the user) if it were non-negative, or -1 if the user desired the system to pick an unused instance number for the newly initiated process. If the user-suggested instance number was currently in use, the process would not be initiated and a pvm error would be returned. 6. Consider, or suggest installing pvmd in a more permanent place. I have pmvd installed in /usr/etc, where a number of other system daemons reside. This has a number of advantages: * It does not get purged by aggressive attempts at clobbering tmp. * /usr/etc is a place where a *single* copy can be shared by multiple machines. If an update is necessary, only one copy need be updated. Updating dozens of tmp directories is unexciting. * It is a location that is guaranteed to be shared only by machines of the same architecture. It has a single disadvantage: * Joe user can't install the system (unmodified) without root access. He could install it locally, or in /tmp, however. ------- I enjoy the system, and hope it evolves. Regards, Duane ------------------------------------------------------------------------------- From: glennw@glenrowan.acci.com.au (Glenn Wightwick) > > Are you using PVM today? Yes > > What are your application(s)? Meteorology and seismic processing. > > How many and what kind of machines have you used with PVM? Sixteen RISC System/6000 Model 350s Cheers, Glenn ------------------------------------------------------------------------------- From: Eric Matthews 1) Yes I am currently using PVM. I just finished a NSF-REU research project in which PVM was used. I also plan to use PVM this fall for an independent study class I will be taking. 2) My applications are mainly computer graphics. The research project I just finished was an object oriented test bed for parallel ray tracing, called Photosynthesis. Future work will also utilize PVM as a platform for other image synthesis techniques. 3) Here at Hope College our computer science department consists of 14 SUN spactstations ( 1, 1+, 2, ipc ) and two Sparc servers. We also have four transputer boards each with eight transputers. It would be nice if PVM could be compiled for use with them as well. Please keep me posted of any changes and additions to PVM. I found it to be an excellent package. And will continue to use it in new projects where there is a need for more computational power. ------------------------------------------------------------------------------- From: whitley@SLCS.SLB.COM > > Are you using PVM today? Yes. > > What are your application(s)? Distributed tape io Distributed operators console Convolutions We are currently attempting to design and implement a concurrent seismic processing system. The actual initial target for this system is a CM5. However, in an attempt to maintain a reasonable amount of portability, we have decide to introduce an application interface on top of CMMD. To test this portability, it is our intention to use PVM to simulate this parallelism on multi-headed convexs and across a network of Sun's. > > How many and what kind of machines have you used with PVM? 20 Sun's 1 Convex 1 cm200 1 cm5 ------------------------------------------------------------------------------- From: Pedro A.M. Vazquez > > Are you using PVM today? > Yes > What are your application(s)? > Fortran o Molecular mechanics and related matrix and force calculations o Quantum chemistry, paralelization of integral calculation C o Molecular Graphics PEX system communicating with the applications above. > How many and what kind of machines have you used with PVM? > Sun4 8 RIOS 8 PMAX 1 I'm trying to port pvm to 386bsd0.1 too. Pedro -- ============================================================================= Pedro A. M. Vazquez | Depto de Fisico Quimica | Instituto de Quimica Universidade Estadual de Campinas | Fone : 39-7253 | Fax : 39-3805 ============================================================================= ------------------------------------------------------------------------------- From: Minesh Amin > Are you using PVM today? Yes, since June 20th '92. > What are your application(s)? > How many and what kind of machines have you used with PVM? Implementing parallel join algos. on cluster of workstations (HP700's and HP300's) Comments, Suggestions := Many "assumptions" are not stated in the doc. provided, examples include i. assumption about the working directory for slaves. ii. the fact that message stack is operated in FIFO iii. the fact that if m bytes are put into the message stack, the size of message received on the other side is NOT m bytes, but (m + ). Same applies to other put[types]. iv. when , say, HP300's and HP700's are used together, whether format of message is checked at runtime or not, and if so NO message is printed at compile/link time ( because machines may be "different", and yet have the same message format ). v. stderr is used as a means of communication when initiating slaves, so that when I have my own debug statements on the slave code, this step is messed up. vi. Only one point, though it may be out of date especially with new releases (I have worked with only upto 2.4.X). In anycase, I don't see the need for "probe" call resulting in a call made to the global pvmd (which may be on a different machine) because if a message is received, the info. is avaiable locally (though I don't know the format). Finally, I would like to say that even with these "problems", I have enjoyed programming using pvmd, and intend using it in the future. -Minesh amin@tera.cs.umn.edu ------------------------------------------------------------------------------- From: fred@wuhydra.wustl.edu (Fred Rosenberger) Are you using PVM today? Not this exact day but am still using PVM. Expect to have an iPSC/860 delivered any day and try on it. We have not actually used PVM for any purpose other than to test its operation yet. Expect to eventually but not this month, or probably next. What are your application(s)? Applications are in quantitative imaging, PET image reconstruction etc. How many and what kind of machines have you used with PVM? Have used with 8-processor SGI, 3 DEC5000 workstations, 4-processor Sun 670MP Our applications have high communication to compute ratio. Using more than about 4 or so workstations can be counterproductive. More efficient protocols for communication would help. We are presently using Ethernet but have performed some tests using FDDI which gives some but not great performance improvment. Ethernet rates are limited by other traffic, FDDI is apparently limited by processor speed to not much better than a byte per us (using test program supplied with PVM). Not having to deal with byte-swapping is great. Any other questions, feel free to e-mail. Fred Fred U. Rosenberger fred@wuibc.wustl.EDU Biomedical Computer Laboratory (314) 362-3124 Washington University 700 S. Euclid Ave. St. Louis, MO 63110 ------------------------------------------------------------------------------- From: koebbe@avatar.math.usu.edu (Joe Koebbe) We have installed pvm but not had a chance to use it. Some time in the next month or so I believe it will receive its first workout. The types of problems that this will be applied are in 1. reservoir simulations; hydrology and petroleum applications 2. laser/optical applications and 3. possibly in some cluster algorithm development. There are probably some other applications, but none I can think of right off. We learned of pvm through the people at Univ of Wyoming (R.E. Ewing and Co.) and will be using it on a system of about 20 Sun workstations. One other possible application I can see is that we may use it for some systems administrative type work. This is rather unclear right now. Joe Koebbe ------------------------------------------------------------------------------- From: "Dr. Edmund West" >Are you using PVM today? We implemented some test applications, and are presently helping two University of Toronto research groups bring up real applications. >What are your application(s)? (1) Test: password cracking program to discover badly chosen passwords. The work was divided among many systems. (2) Astrophysics: hydrogen molecule collisions in interstellar space. This is in the early stages of being converted from sequential programming. (3) Management Studies: min/max problem in production line fabrication. This code is just about to be converted. >How many and what kind of machines have you used with PVM? We have used as many as 29 systems in one of our tests. They include our own DECstation 2100, DECstation 3100, IRIS 4D/70GT, Sun-3/280, as well as DECstation 5000, Sun-4, and IRISs which belong to other groups. ------------------------------------------------------------------------------- From: tim@osiris.usi.utah.edu (Tim Burns) I am using PVM and HeNCE as a student and have written a basic LU decomposition and several BLAS in HeNCE. I will be taking a 500 level linear algebra class in the fall, and hope to learn methods of doing things like QR factorizations in parallel well enough to apply them to PVM or HeNCE. I have noticed that HeNCE doesn't always parallelize code correctly. It would be nice if it was closer to PVM. I also work for a computing facility that plans to install PVM on a cluster of IBM RS/6000 560s and do intesive computing using that facility. Tim Burns ------------------------------------------------------------------------------- From: windemut@lisboa.ks.uiuc.edu (Andreas Windemuth) Here are some comments, dating back half a year, might not be up to date. > Are you using PVM today? Not currently (thesis), but intend to use further. > What are your application(s)? Molecular dynamics > How many and what kind of machines have you used with PVM? approximately 10 NeXTs approximately 20 SGIs > Any comments and suggestions for improvement are welcome. Some communication delays were encountered, Network Linda seemed to be faster. Just an informal observation, no solid evidence available. PVM does not run in parallel on parallel machines, like the iPSC860. Would be nice if it did. Most wanted: Implementations for CM-5 and Parsytec GC. --- Andreas Windemuth +-------------------------------------------------------------------- |Theoretical Biophysics windemut@lisboa.ks.uiuc.edu |University of Illinois Tel: (217)-244-1612 |3121 Beckman Institute Fax: (217)-244-8371 |405 N Mathews, Urbana, IL61801 NeXTmail Ok +-------------------------------------------------------------------- ------------------------------------------------------------------------------- From: lovell@cri.msfc.nasa.gov (Ron Lovell) Here are replies to questions sent 08/11/92: > Are you using PVM today? No, not yet. We do plan to do so this Fall. > What are your application(s)? As part of the system staff, we will experiment with PVM so we will be in a position to recommend and support it for the NASA/Marshall user community. Their applications are numerous. > How many and what kind of machines have you used with PVM? None. Initially we will use old Sun3 workstations for experiments. Possible target systems include CRAY, SGI, RS/6000, SPARC, and DEC RISC, among others. ------------------------------------------------------------------------------- From: ph@mema.ucl.ac.be (Peter Henriksen) We are mailing our full report concerning our use of PVM. Peter Henriksen, Roland Keunings ------------------------------------------------------------------------------- From: calkinsc@tardis.union.edu (Charles Calkins) We use PVM in an academic capacity here at Union College. I had obtained and installed PVM for experimentation with a type of concurrent programming system. My master's thesis mainly focused on a concurrent program that used RPC for communication, but I also showed samples of other methods, including PVM. My PVM application was basically similar to your Mandelbrot sample, although it used the distance estimator method of computation, and displayed its output in an X window as it was computed. We have a fairly small network of Sun workstations and DECsystem 5000s, but up to eight CPUs were used for this application, producing a nearly linear speedup. Each row would be sent to be computed by a worker process, and when the worker completed and returned the calculated line, it would be displayed in the X window. Visually, one may see an empty line between filled in lines for a few seconds until it was finally completed - the worker for the missing line had not returned its result before the worker for the next line did - this allowed one to visually see the concurrent nature of the computation. PVM is currently being used in an independent project by another graduate student, although I do not know the nature of the project myself. Please contact Prof. David Hemmendinger, hemmendd@athena.union.edu, his project advisor, for details. I do wish to say I like the transparency of the PVM system very much. Being able to handle different machine architectures and different byte orderings is a very good feature. I have also used Synchronizing Resources (SR) v1.1 for another project, and version 1.1 did not posess this capability, restricting it to a collection of sun3s. Charles -- Charles Calkins, Union College CS Department ex-Graduate Assistant Internet : calkinsc@tardis.union.edu Starting end of August: Concurrent Systems Group 515 Jolley Hall Washington University St. Louis, MO 63130 ------------------------------------------------------------------------------- From: porter!dem@csn.org (Dennis McNally) We are not currently using the PVM system but a group of out colleques at SSESCO (ssesco.com) are porting an application which we hope to use. We have run only small tests of PVM on our network consisting of a SPARC1, a SPARC ELC and an RS/6000 530. The PVM system is very impressive, if only we had more time. Keep up the good work, dennis mcnally senior atmospheric scientist ------------------------------------------------------------------------------- From: aburke@cse.ogi.edu (Andrew Burke) I am not using PVM today, although I did sent it to a potential customer who might be using it. They wanted our product but ours isn't ready, so I sent them yours. I realize yours is not a commerical product and they realize this as well. They are doing some oil exploration stuff, dunno what. They were using on an IBM RS/6000. Thanks for the system, hope it helps them andrew burke aburke@cse.ogi.edu ------------------------------------------------------------------------------- From: rgb@phy.duke.edu (Robert G. Brown) I have pvm downloaded, and fully intend to use it on as many as 100 workstations on campus (which should give me several hundred MFLOPs in my application) but I have not yet actually used it. My intended application(s) are all various coarse-grained simulations on lattices of varying dimension, typically with nearest-neighbor coupling. Both ODE solution and ordinary simulation techniques are involved. Initially I will probably use pvm to run these problems on a strictly local (single IP subnet) network in the physics department here. This gives me immediate access to 20-30 Suns (varying from Sparc 1 to Sparc 2 in capability, but with at least some Viking based machines in the near future). If I am successful in a port to NeXT's, we also have around 15 NeXTstations. From there I intend to a) add processors locally and b) branch out into some of the unused CPU around campus. I estimate that there are at least 100 1-4 MFLOP machines that are 90% idle 90% of the time in the combined physics, computer science, engineering, chemistry and math departments. If I can tap into a reasonable fraction of this unused potential, I can achieve the throughput of a significant vector machine or better without much difficulty. Long range plans (assuming pvm works as advertised and is scalable to the needs of my problem) include adding a bank of (say) 4-processor Sparcstation 10's, tightly coupled via FDDI and gateway'd onto the physics network via one or two front end (high resolution graphics and mass storage) engines, probably single processor SS10's with 24-bit color and a disk farm. At local University prices, I could in principle construct a completely scalable simulation engine capable of roughly 80 MFLOPs/chassis in ordinary, non-optimized, (but parallelized) code for maybe $40K/chassis. $250K would then buy over 400 MFLOPs and the front ends, and the entire system could be upgraded in place by either modernizing the processors or adding chassis', both at modest cost (compared with the cost of throwing a $12M Cray Y/MP away). I am an experienced systems administrator and can construct and run the whole thing without additional help on a few kilowatts of ordinary electricity (compared to the rather large staff and physical installation required to feed and groom a Cray). With FDDI I hope to achieve high enough machine-to-machine bandwidth to enable medium-grained problems to be attacked as well. Needless to say, I am relying heavily on pvm (or if not the current pvm, a successor) being an efficient vehicle for the construction of cheap, scalable, parallel supercomputers out of over-the-counter stock. If you want suggestions for the future, you might work on per-architecture "specials" that would enable pvm code to use the local Mbus channels on the multiprocessing Sparcs AND the FDDI channels in particularly efficient ways. The Mbus is (as I recall) 300 Mbs vs maybe 100 Mbs for the FDDI channel, so this would affect how problems are divided up. Similar custom mods might work for other multiprocessing architectures (like SGI's). rgb Dr. Robert G. Brown System Administrator Duke University Physics Dept. Durham, NC 27706 (919)-660-2567 Fax (24hr) (919)-660-2525 rgb@phy.duke.edu rgb@physics.phy.duke.edu ------------------------------------------------------------------------------- From: S.C.Lan Yeah, I am now using PVM. I have installed PVM on IBM RS6000/560. Since we have now four IBM RS 6000/560 in our center, we plan to develop some applications on them. Our center is called "National Center for High-Performance Computing". This is the only one Super-computing center in Taiwan. We will try to develop some scientific computing applications in the near future. As to my personal interest, I have try to solve a typical Computational Fluid Dynamic equation, imcompressible Navier-Stoke equation, on HP/NCS. I plan to port my program to PVM to get the preliminary experience on PVM and IBM RS6000. S-C Lan, National Center for High-Performance Computing, Taiwan ------------------------------------------------------------------------------- From: subramon@kazoo.Berkeley.EDU (Ramesh Subramonian) Are you using PVM today? Yes. What are your application(s)? Computational Chemistry. How many and what kind of machines have you used with PVM? 16 SPARCstation2-s over an Ethernet Any comments and suggestions for improvement are welcome. Broadcast is slow. A reduce (or gather) function would be helpful. ------------------------------------------------------------------------------- From: Aad van der Steen 1) We are about to install PVM (so we are not using it yet). 2) We intend to use it on some workstations (2--3 Suns, a SGI Indigo, a Convex 220, and, if possible, our 4-node Meiko i860 system. 3) From other users I have heard that debugging applications executed with PVM is quite hard at the moment (again, we don not have actual experience ourselves yet) because of lack of a good tracing facility. I realize that providing such a tool is not easy, but it would be extremely useful. So, it would be highly appreciated. Best of luck with the further development. Best Regards, Aad van der Steen -- ------------------------------------------------------------------------------- From: Sami Saarinen > > Are you using PVM today? Yes. > What are your application(s)? Today: - building jacket routines above PVM to accomplish automatic error checking and to make library routines easier to use in case of master-slave paradigm. - adding shared memory management (into machines where it's possible) and investigating packing methods in order to reduce data transfer. In case of shared memory, fully remove large data transfer if processes are created by pvmd are into the same (multiprocessor, shared memory) machine where data were supposed to transfer. - Tuning molecular dynamics code to run effectively under PVM. Load balancing will be included. Later: - Parallelizing FEM-code in structural analysis to take advantage of PVM. > How many and what kind of machines have you used with PVM? > We have installed PVM to all of our publicly accessed machines. These include: - Cray X-MP EA/432 - Convex C3840 - Convex C220 - Silicon Graphics 4D/380 VGX - SGI 4D/35 (1), SGI 4D/2x's (several, < 5), SGI Indigos (several, < 5) - Sun 4/390S (will be replaced soon by SUN Sparcserver 690, 2-procs) - SUN Sparc IPX (at least 1) - SUN3 (several, but will be soon upgraded/obsoleted) > Any comments and suggestions for improvement are welcome. > Please send the information to pvm@msr.epm.ornl.gov. > Your help is much appreciated. > Ok. Thanks 4 u, 2, Sami +---------------------------------------+------------------------------------+ ! Sami Saarinen ! E-mail: sbs@csc.fi ! ! Principal Analyst (MSc in Mech.Eng.) ! Sami.Saarinen@csc.fi ! ! Centre for Scientific Computing (CSC) ! Phone: +358 -0- 457 2713 (Int) ! ! P.O. Box 40, Tietotie 6 ! 90- 457 2713 (Nat) ! ! SF-02101 Espoo ! FAX: +358 -0- 457 2302 or ! ! Finland ! 464 803 ! +---------------------------------------+------------------------------------+ ! FEM (Structural Analysis), Benchmarking, C++/C/Fortran, Distributed Comp...! +---------------------------------------+------------------------------------+ ------------------------------------------------------------------------------- From: Rainer Sinkwitz > > Are you using PVM today? YES > > What are your application(s)? Real-time particle simulation, see my talk at HNBCC '91 at fsu. > > How many and what kind of machines have you used with PVM? Up to 25. But since Workstations get shut down here without warning I normally use about 7. > > Any comments and suggestions for improvement are welcome. - PVM should be able to survive workstation failures and shutdowns. oooooo oooo Rainer Sinkwitz, Multimedia Lab sinkwitz@ifi.unizh.ch $ $ $ " University of Zurich, Switzerland VOICE +41-1-257-4346 $"$$ """"$ Inst.f.Informatik, Winterthurerstr. 190 FAX +41-1-363-00-35 o$o "$o $ooo" CH-8057-Z\"urich ------------------------------------------------------------------------------- From: Wierse@rus.uni-stuttgart.de (Andreas Wierse) | | Are you using PVM today? Yes, but it's more like that we are starting to use PVM now. We made some test, whether it works :-) and how fast it is, how easy it is to use, etc. and I think we will take it as a base for a distributed visualization system we are going to develop in the context of a european community project. We hope that it will fit our needs in respect to speed. | | What are your application(s)? our application is a kind of distributed visualization system integrated with a collaborative working environment to allow scientists at different sites, connected with high speed networks to work together on the same apllication | | How many and what kind of machines have you used with PVM? up to now (test phase) I only had five machines connected together with PVM. This won't be exceeded by the project as far as I can see it now. These machines are a CRAY-2, a CRAY-YMP, A Silicon Graphics 340VGX, 310 VGX and an Indigo. A machine that might be included in the future is an Intel Parallel Computer. | | Any comments and suggestions for improvement are welcome. I was quite impressed by the ease of installation and use (I didn't expect it to work right out of the box (or tar-file :-). I just had a few problems due to misunderstandings on my side and one typo (caused by me of course :-). We were very surprised when we tried to find the options mentioned in the documentation or articles that come with PVM: we didn't find any shared memory routines or routines for resource locking. It would be good, if You could keep the documentation up to date (shared memory went out after version 1.0 and now we have 2.4.1 !!). I was a little bit confusing, especially because we need shared memory management a lot. And distributed virtual shared memory would have been quite helpful for us. -- Andreas Wierse | Institute for Computerapplications II | Dep. Computersimulation and Visualization wierse@rus.uni-stuttgart.de | Computer Center University of Stuttgart Tel.: ++49-711-685-5796 | Allmandring 30 Fax: ++49-711-682357 | D-7000 Stuttgart 80 Germany ------------------------------------------------------------------------------- From: Maurits.Malfait@cs.kuleuven.ac.be (Maurits Malfait) > Are you using PVM today? Yes. > > What are your application(s)? Medical image enhancement with a (time consuming) Monte Carlo algorithm. We are experimenting with PVM because we need speedup, but cannot use our parallel computers at computer science dept. for daily use in the hospital. If we manage to develop a parallel algorithm and can run it somewhat efficiently on a cluster of workstations in the hospital, then the people there (first engineers, later medical doctors) would be very pleased, we hope. > > How many and what kind of machines have you used with PVM? Never more than 6 machines simultaneously, because currently we use a master slave model, where communication over the Ethernet easily becomes a bottleneck for slave to master communication. Machines: DEC 3100 and 5100, Sun 4, IBM RS6000. Main difference wrt parallel homogeneous environments is the difference in processor speed which spoils load balance. > We are very pleased with PVM because it is easy to install and start up; and it was not difficult either to port our iPSC code to PVM. We have not used HENCE so far, but will probably do so, because we suppose it might help to monitor and debug. As a suggestion: I think that PVM can gain widespread popularity if you manage to keep it as simple to install and use as it is today, or even can improve this. Not only computer scientists should use it, but also medical doctors, or researchers in a chemistry department; to mention two examples that I am aware of. Hope this is of some use for you. Greetings, Maurits -- +-----------------------------------+--------------------------------+--------+ | Email: na.malfait@na-net.ornl.gov | Maurits Malfait | K KU U | | or, preferably | Katholieke Universiteit Leuven | K KU U | | Maurits.Malfait@cs.kuleuven.ac.be | Department of Computer Science | KK U U | | Phone: + 32 16 201015 x3080 | Celestijnenlaan 200A | K KUUU | | Fax: + 32 16 205308 | B-3001 Leuven (Belgium-Europe) | LEUVEN | +-----------------------------------+--------------------------------+--------+ ------------------------------------------------------------------------------- From: jf@edv6000.tuwien.ac.at (Josef Fritscher (58801-5505)) > > Are you using PVM today? Yes, it's Version 2.4.1pl1 and we will very soon run production code. > > What are your application(s)? Weinbergers KKR-CPA code, Terrain-Correction via FFT, Molecular Dynamics (Biomolecular Simulation), Comparison of Parallel Programming Tool, Parallel Lanuguage Design. > > How many and what kind of machines have you used with PVM? > nine RS/6000-320H configured as parallel machine, and for test purposes a network of RS/6000, DECstations, Suns, and PC's and a Sequent Symmetry with 20 processors. > Any comments and suggestions for improvement are welcome. I'm writing on a Technical Report about PVM and HeNCE and related work. An important one: the release of new versions should find its way to public more generally (e.g. the PVM Mailing list, like Adam Beguelin has done in the NetNews Group comp.parallel with Patch 1. -- Josef Fritscher Internet: fritscher@edvz.tuwien.ac.at Computing Center Bitnet: fritscher@awituw64.bitnet Technical University of Vienna in reality: jf@edv6000.tuwien.ac.at Wiedner Hauptstrasse 8-10 Voice: ++431 58 801/5505 A-1040 Vienna, Austria/Europe Fax: ++431 58 74 211 Member of the Austrian Center for Parallel Computation (ACPC) ------------------------------------------------------------------------------- From: vincent@pastek.cray.com (Vincent LE BARAZER) We are in the progress of evaluating PVM in France , but more tests has been achieved in Cray Inc . We are still using the application examples provided in the distribution package. We run tests all over our Sun SPARC network ( almost 50 cpus ! ) and soon between our 15 Cray systems ( almost 100 Cray cpus ! ) . Regards, Vincent Le Barazer Cray Research France ------------------------------------------------------------------------------- From: tbel@parcb.cern.ch (Tim Bell) 1. Yes, we are using PVM today. 2. We are converting a set of high energy physics detector simulators to PVM. 3. We are using a cluster of 8 IBM RS/6000s model 550s. They are connected in groups of 3 using the serial optical connections. We are using both PVM 2.4 and the PVMe which is customised for the RS/6000 and serial links. The problems we have encountered so far as related to three areas. a. The debugging environment is very difficult. We need to run a debugger on each system to try and attach to executing process. b. There needs to be a set of more sophisticated functions built on top of PVM to perform distributed I/O, gridding, request/gather broadcasts etc. I have not been able to find libraries to do this anywhere. c. Hence is very limited in the problems that it can tackle. The restriction of Acyclic graphs only has meant that we cannot use it for our applications. Tim. ------------------------------------------------------------------------------- From: schreide@ockham.physik.uni-mainz.de \documentstyle[12pt]{letter} \begin{document} \signature{Gernot Schreider} \begin{letter}{PVM--team} \opening{ Dear colleagues of the PVM--team,} Yes I am currently using the PVM-system. I am a theoretical physicist and currently I am doing my PhD on spinglasses. I have to deal a lot with combinatorial and graph-theoretical problems. Among many other stuff I have to compute the weak lattice constants of a lot of graphs (whatever this may be, it's a combinatorial problem though). This problem is completely separable for each graph, so I am using a very simple master--slave concept for distributing all graphs over three IBM RISC/6000 workstations. The data transfer is just a few bytes, namely the description of the graph being sent to the slave and the result, which is just an integer, is sent back. Nevertheless I did some timing on data transfer via TCP/IP on ethernet. You may find this interesting and I would like to hear some comment about these numbers I have figured out. Timing is not so easily done, one must not use the processtime on a many user machine. So I measured the real time for sending all data and receiving an acknowledgement signal. \newpage \begin{center} { \Large performance of datatransmission via TCP/IP on ethernet } \end{center} {\bf variable datapackagesize with fixed overallsize of 10 MByte}\par \begin{tabular}{ccc}\hline packagesize & time [sec] & kByte/sec\\ \hline 1 kByte & 17 & 602\\ 10 kByte & 39 & 263\\ 100 kByte & 26 & 393\\ 1 MByte & 31 & 330\\ \end{tabular} {\bf variable number of datapackages with fixed packagesize of 1 kByte}\par \begin{tabular}{ccc}\hline number of packages & time [sec] kByte/sec\\ \hline 10 & $<$1 & -- \\ 100 & 2 & 50 \\ 1024 & 17 & 60 \\ 10240 & 167 & 61 \\ \end{tabular} {\bf variable number of datapackages with fixed packagesize of 1 kByte}\par \begin{tabular}{ccc}\hline number of packages & time [sec] kByte/sec\\ \hline 10 & 1 & 1000 \\ 100 & 4 & 250 \\ 1024 & 39 & 262 \\ 10240 & 399 & 257 \\ \end{tabular} {\bf variable number of datapackages with fixed packagesize of 1 kByte}\par \begin{tabular}{ccc}\hline number of packages & time [sec] kByte/sec\\ \hline 10 & 3 & 333 \\ 100 & 27 & 370 \\ 1024 & 264 & 388 \\ \end{tabular} \newpage Coming now to the suggestions I want to make, I have to say that working with PVM was easy for me, maybe this is due to the fact that I am not using complicated net structures. As I can remember there was a possibility of using shared memory in earlier implementations. This would be very interesting for me, because I could hold a big list of SAW's in one machine only and needn't set up this list in all slave--'machines'. Please let me know what's actually going on with this facility. The second point is the following, suppose one part of your net went down or one part of your program failed in one instance. Then you fix this problem while the program is still running. Having fixed everything, you want to enlarge your system again by using this particular machine, which is up again, or by enrolling another instance of this special part of the program. This is not so easily done, as far as I know, or am I wrong? Finally I want to say that you are doing very good job. \closing{Sincerely} \end{letter} \end{document} ------------------------------------------------------------------------------- From: Oliver Wendt Dear PVM Project Group, I received your e-mail today and I would like to report on the state of the art of our PVM. We installed it on a SUN cluster ( two SPARCs and six diskless ELCs ) and the pvmd runs fine on all of them. But unfortunately we can't even get the simple mandelbrot example working properly. When started, it states: deep-thought# mandel_main 100 100 0 0 50 50 xxx using 4 tile servers in calculation PVM user lib : initiate: No such File initiate: No such File can't initiate calculator New epoch 1 deep-thought# mandel_main 100 100 0 0 50 50 whole_mandelbrot using 4 tile servers in calculation PVM user lib : initiate: No such File initiate: No such File can't initiate calculator New epoch 2 So if you would kindly help us with that problem, we would indeed like to use your PVM ( mainly for the distribution of genetic algorithms over the net ) and reply to your questions. With best regards Oliver Wendt & Peter Rittgen ------------------------------------------------------------------------------- From: steve@cs.clemson.edu We are considering the PVM package to support courses at Clemson. If we decide to do so, the most likely platforms are our Suns (both 3 and 4). ------------------------------------------------------------------------------- From proctor@cam.enet.dec.com Wed Aug 12 09:11:46 1992 I was looking into using portions of pvm to represent recipes (steps and flow) within manufacturing processes - a perhaps rather unique application. Unfortunately, my efforts have be redirected over the past several months and I have not been able to pursue this application. regards, bob ------------------------------------------------------------------------------- From: galagun@cs.sandia.gov (Glenn A. Laguna) Thanks for the message. I really haven't started using PVM yet, but plan to use it for a Boundary Element CFD code and a Direct Simulation Monte Carlo (DSMC) code. ------------------------------------------------------------------------------- From: fred@wuhydra.wustl.edu (Fred Rosenberger) I responded to your query earlier but have additional comments now. 1. A great help would be more performance information. How fast can I expect to carry out message passing with an up to date system? For example, if I have a choice between a 16 processor iPSC/860 using INTEL software and 12 HP-730 workstations connected over a dedicated FDDI network, which will be faster or more effective? This should be problem dependent with the ipsc/860 winning when the communication requirements are high and the HPs winning when the communication requirements are low. I have characterizations of my problems in terms of computing and communicating requirements. I can't get 12 HP-730s with FDDI, install PVM, and run tests. I can run simple benchmarks on one 730 or use SPECMARKS or some such standard benchmark for comparison to systems on which the code is running now. What I want and think could be supplied is more on predicting PVM communication performance on an HP 730 or other systems. Then I could estimate performance before choosing. Better characterization of communication would also help me decide whether a particular application should be considered for use with PVM at all, or how many processors it makes sense to consider using on a particular application. With our DEC5000 and SPARC2s on a particular problem, it seems that using more than a few processors is counterproductive. It would be good to predict this in advance. Then, if I performed a few measurements, I could compare to prediction to see if I was close or not. If not, I should investigate further to understand why predicition and performance were different. I'm all for performance monitoring but I'm also in favor of performance prediction. We learn different things from each. 2. I have not tried to use hence. If I understand its operation correctly, it would not work for my initial application. I have some routines that initially get a lot of fixed state information and parameters, then are called 200 times with a new input data array each time. It appeared to me that there was no way to do this efficiently (that is without resending the state and parameter information each time). I think I need a node with multiple inputs, one for initialization that is used once, and another for the data to be processed that can be used any number of times. Fred Fred U. Rosenberger fred@wuibc.wustl.EDU Biomedical Computer Laboratory (314) 362-3124 Washington University 700 S. Euclid Ave. St. Louis, MO 63110 ------------------------------------------------------------------------------- From: Perception is deception. DTN 297-6383 Are you using PVM today? > Not really using it yet. We're looking into porting PVM and HeNCE to OSF/1 > on Digital's Alpha architecture. We're an engineering group in Digital's > Education/Science Marketing Business Unit. I know that Vehbi Tasar's group > is also in touch with you, so I should mention that we have different, but > coordinated objectives. Give me a call if you'd like to know more about what > we're up to - my number is 508-467-6383. What are your application(s)? > We'll likely just be testing with supplied examples here. How many and what kind of machines have you used with PVM? > Will be testing with Alpha systems running OSF/1, and MIPS systems running > OSF/1 or ULTRIX. Any comments and suggestions for improvement are welcome. Please send the information to pvm@msr.epm.ornl.gov. Your help is much appreciated. > Our focus is on the current versions, so most likely we will relay any > suggestions through Vehbi Tasar's group. Best wishes, The PVM Project Team > Likewise, > Chuck Schneider > > My coworkers include Andy Vesper, Robert Yuan, and manager Roland Belanger. > Actually, I'm only peripherally involved at this point, Andy and Robert are > doing the work. ------------------------------------------------------------------------------- From: braaten@buffalo.crd.ge.com (Mark E. Braaten) At GE, our interest in PVM is a possible candidate for writing parallel message passing codes that are portable between different parallel computers and also across networks of workstations. We have just recently requested the pvm source from netlib and are attempting to install it on an Intel iPSC/860 machine. This installation process is not as straightforward as we had hoped, and we have encountered what appear to be bugs in the installation procedure. The manner in which the source code arrived, in seven out of order e-mail messages, is quite inconvenient to start with. In the words of one of our programmers, he said it was the first time "he had ordered a car and had in arrive in seven unmarked crates". Attempts at installing pvm for the iPSC/860 result in code that runs on the host machine (386-based), but not on the nodes. The makefiles in the I860 directories use f77 and cc compilers, which are the compilers for the 386, not the if77 and icc compilers required for the i860 nodes. Attempts to compile and link using the icc and if77 compilers fail in the loader step since the -lsocket reference is not satisfied. Our applications of interest include CFD, optimization, Monte-Carlo combustion models, and image processing. We already have parallelized some applications using the Intel message passing libraries. We would like the resulting PVM codes to run on Intel, TMC, future Cray, future CONVEX, future IBM massively parallel machines, as well as on Sun, HP workstations, and conventional Cray and CONVEX supercomputers. Please have someone in your group contact me to discuss the problems we are having installing pvm for the nodes in our iPSC/860. Thanks in advance for your help. Sincerely, Dr. Mark E. Braaten GE Research and Development Center 518-387-6582 braaten@crd.ge.com ------------------------------------------------------------------------------- From: weiming@cc.gatech.edu Are you using PVM today? Yes. What are your application(s)? Simulations. How many and what kind of machines have you used with PVM? SUN SPARC1, SPARC2, Intel iPSC2 Weiming Gu weiming@cc ------------------------------------------------------------------------------- From: mryan@bu-pub.bu.edu (Mark Ryan) > >Are you using PVM today? As the system administrator it was my job to install pvm. I have only used it to the extent of running the supplied demos. I have not used it today and am not sure who is in fact using it. >What are your application(s)? See above. >How many and what kind of machines have you used with PVM? 40+ machines. Sun3 and 4, decstation 3100, 5000, vax, SGI's, CM2, CM5, IBM RT's and RS machines. >Any comments and suggestions for improvement are welcome. a) It would be nice if PVM understood somwthing of machine architectures and load averages. Often one machine can solve a problem much faster than using several. b) Better monitoring tools. (Not hence). It would be nice to see the execution times while running PVM to find out which machines are causing the slow downs. c) It would also be nice to see network stats. d) Building and maintaining executables on every machine is a real pain when working with a lot of machines. Maybe the daemon could determine if an executable for a given architecture is available and make copies to the local host. Comment: Hence is not used do to execution problems. Some of the supplied demos do not work. Seems to need more work. Comment: Thanks for an interesting package. If I used it more I could probably be more constructive in my comments. Keep up the good work. ------------------------------------------------------------------------------- From: Claudio Condini > > Are you using PVM today? Yes. At the moment ours is more of an evaluation than a serious application but if the results are satisfactory then we'll move on to the real application after September. > > What are your application(s)? Interactive processing, scientific computing. We have also done performance evaluation of PVM (especially against ISIS). > > How many and what kind of machines have you used with PVM? We have a cluster of 6 HP-750s. We have also used PVM on IBM RS6000 (3) and SUNs (3 or 4). We would like to try PVM with hundreds of processes running on 25 or 30 HP700s. > > Any comments and suggestions for improvement are welcome. In order of priority, the changes we would like to be made to PVM are: 1. Passing parameters to slave processes (ie in the initiate()) 2. Fault tolerance: if a slave dies we would like the master to be notified. At the moment we can do it by hand but it would be nice to have it as part of the PVM system. We are thinking of extending PVM to embed 1 and 2. Regards, Claudio Condini, Fabien Collin. ------------------------------------------------------------------------------- From: rhollen@valhalla.cs.wright.edu (Richard L. Hollenbach III) > > Are you using PVM today? We are planning to incorporate it into our computational chemisty environment to spread out the computations. It is currently in the development stages, not in production. It is more of a research project than a company issue. > > What are your application(s)? We do molecular modeling. I work for Marion Merrell Dow Pharm. in the Research Institute for Theoretical Chemistry branch in Cincinnati. We run mostly third-party applications, but we have some source code to work with. > > How many and what kind of machines have you used with PVM? We are primarily UNIX based, but we still have a few VAX's in use. We have a Cray Y-MP2E, an SGI 4D/380VGX, an Evans and Sutherland ESV 50 and 20 and a few SUN and DEC UNIX workstations. We have PVM working on all of the above UNIX systems using the Cray and the SGI as compute servers and the SGI and ESVs as graphic platforms. I had to port PVM to the ESVs, but it only took a few mods to the Makefiles. > > Any comments and suggestions for improvement are welcome. I learned about PVM through Cray Research in one of their service bulletins. I then had a class in distributed programming at Wright State Univ. (I'm working on my Masters in Comp Engr) and chose PVM as a research topic for the class and presented a lecture and paper on it. I then decided to continue the work here at MMD. ------------------------------------------------------------------ Richard L. Hollenbach II Mail: rhollen@valhalla.cs.wright.edu CEG Graduate Student Cray System Manager Phone: (513) 948-7027 R&D I/S FAX: (513) 948-7432 Marion Merrell Dow Room: 31-341 ------------------------------------------------------------------ From: "Caplin Cybernetics Corp. Lt" In reply to your questions: 1. YES we are using PVM today. 2. We are not yet using it in anger on "real" applications, although we plan to. We specialize in parallel machines, both networks of IBM RS6000s and our own multiple-Intel i860XP machines (using transputers for communication between nodes). (I suppose this answers question 3 too). We would like to run PVM on our own boxes too, giving us a heterogeneous communications system between standard UNIX nodes and our own go-faster nodes. Whether we do this using TCP/IP, or bypass all that and use our own lower-level message passing to provide the communications backbone is still open to debate. Please keep me posted on PVM advances! Best regards, Paul Williams Head of Software Development Caplin Cybernetics Ltd, London UK. ------------------------------------------------------------------------------- From: mep@inel.gov (Paul Murray) Are you using PVM today? Yes. What are your application(s)? A parallel implementation of the preconditioned conjugate gradient method. How many and what kind of machines have you used with PVM? CRAY, HP9K, PMAX. ------------------------------------------------------------------------------- From: kyle@teak.cray.com (John Kyle) Dale Carstensen of Los Alamos asked me to send you my answers to the following questions. I hope that they prove useful. Please contact me if you have any further questions. -- John >Are you using PVM today? YES >What are your application(s)? A variety of distributed models. Specifically, a coupled ocean-atmosphere model distributed between a Y-MP and a CN2 and others >How many and what kind of machines have you used with PVM? Sun, SGI, Y-MP, CM-2 (also writing codes codes for Cray MPP) So far, I have only run a code on a 12 node metacomputer. >Comments... 1. I find it quite painful to coordinate development of software in a heterogeneous metacomputer environment. It would be nice to have some tools like make which operate in a heterogeneous environment. Also, source code control is made quite difficult in heterogeneous environments without distributed file systems. And, today anyway, distributed file systems do not perform extremely well on supercomputers. < Points 2 and 3 may result from my immaturity with PVM, so please > < forgive any obvious stupidity on my part. > 2. I am currently doing some studies on the the best ways to send multidimensional data from one host to another. Some advice from those who know PVM well would be very welcome. 3. Finally, I would like to able to "sense" the state of another PVM particpant. For example, in my coupled ocean-atmosphere model it would be most useful to be able to tell if the atmosphere has posted a blocking receive. If it has, then the ocean will send data directly across the HiPPI. Otherwise, the ocean would write the data to the DataVault and notify the atmosphere that the data is available. -------------------------------------------------------------------- John W. Kyle Cray Research, Inc. Sr. Computational Scientist 655F Lone Oak Drive, Eagan, MN 55121 MPP Applications (612) 683-5804 FAX: (612) 683-5276 ------------------------------------------------------------------------------- From: David Thomas Status: R Are you using PVM today? Literally today -- No. But I have been using it. What are your application(s)? I'm running a code to calculate the primordial abundances of elements from inhomogeneous models of nucleosynthesis. The calculation is not particularly amenable to parallelization, however, since each calculation takes about half an hour on a Sparc II, and I need to run a couple of hundred data points to get any meaningful results, that's several days of computer time. I've used PVM to simply send a data point to each of several machines, wait for one of them to return, and send it a new set of data. Meanwhile it collects the data that returns, and saves the results. How many and what kind of machines have you used with PVM? Most of the runs are done on three machines - a Sparc II, a Sparc IPC, and another sparc machine (I'm not sure what it's called, but it's a multiprocessor machine). I have at times also run it a cluster of about 9 Decstations (5000/200 and 5000/125). Any comments and suggestions for improvement are welcome. Two comments: It would be useful if I could give an argument to the master pvmd to tell it to "nice" all the slaves. At the moment I have to either run everything at full priority (a bit anti-social on other people's workstations) or "rsh" each machine after I've started and renice all the daemons. The other thing is (though I'm not sure how one would implement it) that the master pvmd has no (convenient) way of telling whether one of the slaves (either the slave daemon, or the node program) has died. It would be useful to be able to recover in such an eventuality. All in all though, I find it a useful piece of software. --- David Thomas University of Chicago davet@oddjob.uchicago.edu Astronomy and Astrophysics ------------------------------------------------------------------------------- From: wrightdv@pwfl.com (David Wright 407-796-8477) Not using PVM today. We are investigating whether or not we will use PVM or stick with other packages. Thanks. Keep up the good work! Dave Wright CFD Computing Support Pratt and Whitney ------------------------------------------------------------------------------- From: gropp@antares.mcs.anl.gov (William Gropp) Are you using PVM today? I'm trying. What are your application(s)? Algorithm development, particularly for PDEs in CFD. How many and what kind of machines have you used with PVM? Suns and Irises. I'd like to use it with something with fast, non-socket oriented communications. Any comments and suggestions for improvement are welcome. Please send the information to pvm@msr.epm.ornl.gov. Your help is much appreciated. Here is a partial list of problems that we encountered in interfacing to pvm: 1) Limited path for looking for executables. 2) Does not set a reasonable working directory. 3) Does not pass command line. 4) There is no routine to return the number of machines of a particular architecture. 5) stderr and stdout go to where the deamons are started up. 6) makes no attempt to catch interupts on user process; one could die and the PVM system wouldn't even know. 7) byte messages are not handled well 8) Setting up the pvm daemon is not as convienent as it could be, and since there isn't a way to query it, any program that expects to work on its own must start its own pvmd. 9) No interface to things like the intel nx communications calls We've fixed 1-3 and 8 (partially) in our interface; 4-7 and 9 require changes to pvm. Bill Gropp ------------------------------------------------------------------------------- From: Kremenek George 1.Are you using PVM today? Yes but only for small test example. Sorry I have no time to do more for now. 2. What are your application(s)? Except test examples none so far. 3. How many and what kind of machines have you used with PVM? 4 SUNs, 1 CRAY Y-MP, 1 HP 9000/720, 2 IBMs RS6000, 1 IBM-PC ( Intel SRM from IPSC860). I would like to run PVM on the NCUBE (128 nodes Mark2) and on a Intel Paragon machine (80- 500 nodes, OSF/1-AD, end of 1992). George --------------------------------------------------------------------------- George J. Kremenek | e-mail: kremenek@sdsc.edu San Diego Supercomputer Center (SDSC) | phone: (619) 534 8324 General Atomics, P.O. Box 85608 | fax: (619) 534 5152 San Diego, CA 92138-5608 P.S. also u9180@y1.sdsc.edu ------------------------------------------------------------------------------- From: dinucci@nas.nasa.gov (David C. DiNucci) I am responsible for acquiring and testing PVM, and making it available to others. As such, the only applications which I have run are the sample applications included with the distribution, and one "embarrassingly parallel" benchmark converted to PVM from iPSC message passing by a colleague (David Browning). I have tested PVM on networks including SGI workstations, Sun SPARC2 workstations, a Cray 2, a Cray XMP, and a Convex. I have the following recommendations to the administrative aspects of PVM: (a) That the "aimk" script be replaced by some other module which finds the current architecture and flags, but which could be called from within other scripts. This would more readily allow the use of this information, not only in an architecture-independent make, but also in installation scripts, etc. (b) That the man page for pvmd not contain the installation procedures, since end users should not need to be aware of these. (c) That the "f2c" library be renamed so that it is identifiable with pvm. Locally, we have renamed it "libpvmf.a". (d) That a man page for the "f2c" library be included with the distribution, and that pointers to this man page be added to the other relevant man pages. I have included our version of the "libpvmf" man page at the bottom of this message. (e) That a "known bugs" portion be added to the pvmd man page. At this time, we have included a small paragraph about the problem that virtual circuits will intermittently fail when system or network loads get too high, evidently due to limited buffering of messages. (f) That the line LIBS = be changed to LIBS = $(PVMLFLAGS) in pvm/chol/makefile (g) That some script, somewhere, create the directory $ARCH under pvm. Without this, the "mv" commands in the test makefiles (e.g. pvm/chol/makefile) move the executables to a new executable named $ARCH, which is incorrect. Overall, I have been fairly happy with the ease with which the package can be installed on several machines, using similar installation procedures for each. Dave ------------------------------------------------------------------------------- From: Geiger@rus.uni-stuttgart.de (Alfred Geiger) here some remarks about the status of PVM at the University of Stuttgart which will hopefully answer your questions. We have just introduced PVM on a cluster of RS/6000 and on 2 Crays for a cooperative research project in parallel and distributed computing between the Aerospace-faculty (especially the people doing reentry- problems for Hermes and combustion) and the IBM science-center in Heidelberg. The RS/6000 cluster is interconnected by IBMs optical link and FDDI and consists of 8 550s. We intend to bring PVM also to another cluster that is in use for the whole university and (a student is just trying to do this) to a MasPar (as atomic resource). Another group here at the computer-center is just evaluating the usability of PVM for their distributed graphics and distributed object-oriented technical database projects. Regards, @lfred geiger ------------------------------------------------------------------------------- From: mhinds@bt-sys.bt.co.uk (Mike Hinds) 1) Yes I am using PVM 2) Dstributing computation for modeling, hope to connect it up to AVS to display the results 3) 12 Sun Sparc2 (on two ethernets with one router) Comments: more example code might me usefull (being on of those programmers who like to see how people use the features offered). whould also be interested in seeing how pvm should be set up to allow different users to share the pvmd's , currently each user starts there own. +---------------------+---------------------+------------------------------+ | Mike Hinds | Parallel and | Software Research Division | | | Distributed Systems | Room 140, B81, | | hinds_m at | as an aid to | British Telecom Laboratories | | bt-web.bt.co.uk | Data Visualisation | Martlesham Heath, | | Tel +44 473 643404. | and | Ipswich, Suffolk | | Fax +44 473 647410. | Virtual Reality | IP5 7RE, UK. | +---------------------+---------------------+------------------------------+ From: Johnny Petersen > > Are you using PVM today? Yes > > What are your application(s)? Air pollution transport, Seismic migration, Visualization, and Ocean Circulation. > > How many and what kind of machines have you used with PVM? Anywhere from 2 to 10 IBM Risc System/6000 raging from 550's to 320's. I am also using the SOCC. > > Any comments and suggestions for improvement are welcome. Please make a better User Manual. I would like to have my programs in a different directory other than /pvm/ARCH. Cheers Johnny Petersen johnny@bsc.no ------------------------------------------------------------------------------- From: duncan@meiko.co.uk (Duncan Roweth) Meiko is close to completion of a second generation product, a distributed memory MIMD machine with sparc processors and optional Fujitsu vector units. The processors are connected via a high bandwidth low latency network supporting user level remote store operations between processes. We are implementing a number of message passing interfaces, our own (CSN) for upwards compatibility for our existing codes, PVM, and others to enable us to port exisiting codes. We'd really prefer that everyone used the same interface of course. We have (minimal) Solaris 2.0 running on the nodes and a TCP implementation running over the comms network. This allows us to run the existing PVM as is. We plan to do a direct implementation, avoiding the socket interface and kernel intervention. This should get message startup latency down to 15-20 microsecs. We would be interested in a summary of responses to your posting if you plan one. Best Wishes Duncan Roweth Meiko Limited 650 Aztec West Bristol BS12 4SD England Phone : +44 454 616171 FAX : +44 454 618188 E-Mail: duncan@meiko.co.uk ------------------------------------------------------------------------------- From: Kazuto Horimatsu We have just started to use PVM, so I can not to tell you much now. We are going to use PVM to solve the linear system of equations which is used for finite element analysis. Machines, which PVM is installed, are Personal IRIS and Sony NEWS now. I am glad that I can use this system. Thank you. ------------------------------------------------------------------------------- From: Michel Gastaldo Are you using PVM? YES What are your applications : 1- Testing PVM on our network 2- Implementing a parallel version of a Monte-Carlo Simulation How many and what kind of machines used with PVM : In a first time 3 Sparc-Stations and 1 Sequent Symmetry (20 processors) In the future : 1 Cray XMP 2 processors and 1 Cray YMP 8 processors Comments and suggestions of improvement : * Improvement of the fiability of the communications (to many problems with the sockets, more specialy on the sequent) * Why can't we make several virtual machines on the same physical machines? In other word, why the intersection of 2 virtual machines has to be empty? * What about the debugging environment? * Couldn't we have an option allowing pvm not to stop when a machine shut-down (useful in the case of a large configuration that does not use all the available machines) * We would like informations about the use of PVM on the future Cray MPP. Gerard Meurant and Michel Gastaldo. e-mail : meurant@etca.fr gastaldo@etca.fr ------------------------------------------------------------------------------- From: Steve Heistand >Are you using PVM today? Sort of, I tried using PVM in conjunction with HeNCE, but for some strange reason HeNCE wouldn't work correctly. >What are your application(s)? I am trying to run some CFD code in parallel, the software normally takes a few days on a Apollo 10000, I would like to speed that up just a bit. >How many and what kind of machines have you used with PVM? At the moment I am using PVM (and HeNCE) with about a dozen Dec stations, (mix ov 3100's and 5000's), but we also have several hundred more dec stations and brand new vax and a bunch of little Apollo's (the convenience of a university) I have given up on HeNCE and am trying to use PVM directly, which looks simple enough, but have not made much progress as of late. I hope this helps. steve ------------------------------------------------------------------------------- From: tremback@janus.atmos.colostate.edu We are just starting to use PVM. Small test programs have been tested and are functioning well. We eventually will use it to MIMD-ize a mesoscale atmospheric model (RAMS). So far, we have used RS/6000 machines in a cluster up to 4. In the near future, we want to use a cluster up to 20 and maybe throw in a STARDENT/Kubota 3000 also. - Craig Tremback ------------------------------------------------------------------------------- From: eddy@duteca.et.tudelft.nl (J.G.E. Olk) > Are you using PVM today? yes. > What are your application(s)? - parallel parsing of formal languages I use PVM here to allow for easy examination of computation times and communication behaviour of the parallel parse algorithm rather than for actually getting a speedup. (The ethernet is definitely a bottleneck here.) - parallel programming excersises for students: - mandelbrot - communication processor For a course on multiprocessor computer architecture students have to write two applications using PVM. The first one is simply parallelizing the sequential mandelbrot algorithm. This is to get them used to PVM and writing (and debugging!) parallel programs. The second exercise is to write a simple communication processor CP and construct a communication mesh consisting of nodes having a CP and a DP (PWM component executing the application code, e.g. mandelbrot). Communication between the PVM components is now done via the network of CPs and not directly by PVM. This is of course an inefficient implementation but the goal of this excersise is to examine the communication behavior, i.e. link usage, buffer usage, etc. > How many and what kind of machines have you used with PVM? 4 Decstations 2100 running Ultrix 3.1 4 HP9000/710 running HPUX 8.07 5 SUN4 sparcs (SLC, IPC) running SUNOS 4.1 / SUNOS 4.1.1 Mostly, the PVM applications were using only one architecture at a time. Students for the architecture course are working on the HP workstations. > Any comments and suggestions for improvement are welcome. > Please send the information to pvm@msr.epm.ornl.gov. > Your help is much appreciated. I would like to see a more detailed user guide, e.g. with an application example and examples on the use of each PVM function. Also a list of PVM errors and their possible causes, e.g. "internal error, should not happen, most likely you are overwriting your own data segment...etc", would be nice. Perhaps the user guide could also contain more information on the 'internals' of PVM, e.g. "PVM daemon uses buffers of x Kb so the maximum amount of data that can be send at a time is about y bytes". I hope this information is of use to the PVM Project Team. Greetings, - Eddy \ Eddy Olk Section Computer Architecture \ / email: eddy@duteca.et.tudelft.nl Department of Electrical Engineering / \ phone: +31-15-785021 Delft University of Technology \ / The Netherlands / ------------------------------------------------------------------------------- From: han@sci.brooklyn.cuny.edu (Mr. Han (McAloon)) I am using PVM, and mostly using it to implement some software parallel application for testing purposes. PVM is a very nice distribute system as comparison to other systems like Linda. Its realiability and speed are very good. It is easy to design the applications. Here, may I give some suggestions for improvement (1)For some applications, I think asynchronous communication is very valuable. So such communication functions will make PVM more enticing. Could you supply such communication and also let users call some function to decide whether they need such services or not? (2)For snd(), there is function probe() to let a process inquire a message without blocking, could there be a similar function for vsnd()? (3)For some programing purpose, we try to fork a child process to do some things in PVM environment, but the child process will not get the same PVM ID as its parent. If the child process gets the same PVM ID as its parent and both processes will receive the same message , and let them decide respond to the message according to being the parent or child. This feature will also very valuable. In the test, using the current version PVM, we get confusion when we send the message to the PVM ID, it is not quite clear whether the parent gets it or the child. Meanwhile, is it necessary to let all slave processes call function whoami() before they start to work in PVM environment? From testing, it sounds that every child process should call function whoami() in order to become a PVM process even though the master process calls initiateM() to generate these slave prcocesses. (4)For each user to run a PVM program, he should start his a pvmd daemon on each machine. Could we start only one pvmd on each machine and let all the users use the same group of pvmds to do their programs? (5)By the way, we use HeNCE, there are some confusion about using the cost matrix to assign different subroutines to different machines. According to the menual, the cell with nonegative cost will let the machine run the subroutines. But from testing, it is always the machine with a blank cost cell running the subroutine; on the other hand, if the column for one subroutine is blank all the way down, HeNCE will complain it. The cost matrix matching mechanism makes a little confusion when one subroutine is supposed to run by one machine while there are several machines without noblank cost for this subroutines if we assume the mechanism works well as the menual mentions. (6)Another interesting question, if I try to write the PVM codes in HeNCE subroutines to let different subroutines communicate during their running. From the pvm console, we can see each subroutine(or each HeNCE graph node) runs as one process on each machine, but when we call whoami() in that subroutine, it just shows the instance number without process name which means we can not get the same PVM ID in the subroutine as the console shows. If we can, then that will allow programmers to write PVM codes to communicate inside subroutines with an additional choice to use NODE interfaces, and HeNCE graphic facility will be avaiable to run any kinds of pvm programs. Thanks. Sincerely, Dongguang Han ------------------------------------------------------------------------------- From: Ming-Jen Chan pvm> Are you using PVM today? Yes. pvm> What are your application(s)? No specific application. Instead, I'm working on a new interface on Mach operatiing systems which allows PVM and applications running on to work in more effcient way. pvm> How many and what kind of machines have you used with PVM? My configuration includes Suns running either SunOS or Mach operation system and PMAX running Mach. Ming-Jen Chan Computer Science, Carnegie Mellon University E-mail: mjchan+@cs.cmu.edu Phone: (W) 412 268-3054 (H) 412 422-8040 ------------------------------------------------------------------------------- From: oli@inel.gov (Mark Oliver) Yep, we're still using it. I've put together applications that involve elastic wave equations and heat transfer. I've taken a fully implicit set of equations and distributed the matrix solve among a collection that includes a CRAY X/MP, several SGI's, HP's, DEC's, and SUN's. The DEC's and SUN's do not appear to be the best platforms as far as network communication is concerned. We've had some problems with the CRAY in the cluster. None of the daemons appear to be able to find the CRAY unless the pvm cluster is started from the CRAY (I think this is a local problem though). The simulator has all my compute slaves push their results for each time step to an SGI and I visualize interactively there. It's pretty neat being able to compute all over the site and still be able to interact with the model through the graphics. Unless, you place the graphics daemon in the foreground with the GL command foreground you need to have the SGI initiate all the primary session or the other jobs can't find the graphics processor. All in all very robust, I'm very happy. Is there anything that your user community can do to help? Visualization? More apps? Mark o o o o o o o .... ________________________ _____=======___________ o _____ ||Mark Oliver | | oli@bambam.inel.gov | .][__n_n_|DD[ ====____ |Scientific Computing | |(208) 526-9952 | >(________|__|_[_INEL___]_|EG&G Idaho, Inc.______|_|_FAX:_526-9908________|_ _/oo OOOOO oo` ooo ooo 'o^o^o o^o^o` ' o^o^o` o^o^o` -+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+- Idaho National Engineering Laboratory, Idaho Falls, Idaho, 83415. ------------------------------------------------------------------------------- From: SCHNEIDER@EMBL-Heidelberg.DE 1 ) Yes I just started to port my program using PVM 2.) Molecular biology and biotechnology application which scans databases of protein sequences for similarities that are significant in terms of protein structures or functions. Has to do with genome projects, Grand challenges... The program is running on a Parsytec machine under Parix. A port to a CM5 and KSR-1 is planned. 3.) because I haven't finished the port: these are the machines I can (will ?) use: up to 10 SPARCstations, different kinds of SGI's, lots of DECstations. A colleague is using NEXT machines. A debugger would be great, Load balancing according to the actual load of a machine would be nice I like it, I think it's very well thought (compared to other stuff I saw/use) Reinhard Schneider European Molecular Biology Laboratory, Heidelberg, Germany Protein Design Group Internet: schneider@embl-heidelberg.de ------------------------------------------------------------------------------- From: koonseng@iss.nus.sg (lim koon seng) > Are you using PVM today? Yes > What are your application(s)? Distributed Simulation. > How many and what kind of machines have you used with PVM? Sun Sparcs 1+ and 2s ------------------------------------------------------------------------------- From: Robert Drees > Are you using PVM today? Yes, but for evaluation purposes only. > What are your application(s)? Parallel Genetic Algorithms > How many and what kind of machines have you used with PVM? About 20 HP 9000 Workstations (300, 400, 700) within a diskless cluster. > Any comments and suggestions for improvement are welcome. 1) Because we have a diskless workstation cluster there was some trouble with the '/tmp/pvmd.' files created by the PVM daemons - there is only one '/tmp' directory for all machines. We solved this problem using 'context dependent files'. Maybe you could mention this problem within the documentation of PVM or, a much better way, change the naming of the files. 2) The behavior of the master PVM daemon to shut down the whole PVM system if any of the slave daemons is killed is a bit annoying. Especially if you want to build a fault tolerant system. Adding and deleting hosts dynamically would be a nice feature. 4) The load balancing between very fast and very slow machines works poor. For example: processes to distribute over 4 Machines (1x 700, 2x 400, 1x 300) distribution: 700 2 processes finished after 20 seconds (by PVM) 400 2 processes finished after 40 seconds 400 2 processes finished after 40 seconds 300 2 processes finished after 100 seconds --------------------------------------------- total time 100 seconds distribution: 700 4 processes finished after 40 seconds (better) 400 2 processes finished after 40 seconds 400 2 processes finished after 40 seconds 300 0 processes finished after 0 seconds --------------------------------------------- total time 40 seconds Thank you for the request and good luck with your project Robert Drees ------------------------------------------------------------------------------- From: dwf@acl.lanl.gov (David W. Forslund) > > Are you using PVM today? Yes > > What are your application(s)? Primarily as a simple replacement for sockets to control a raw HIPPI channel between a Cray and the CM-2 > > How many and what kind of machines have you used with PVM? Sun's, SGI, and Crays. > > Any comments and suggestions for improvement are welcome. Event-driven simulation model is needed. User-specified buffers should be allowed to minimize copying. Multi-network connection capability is important. HeNCE needs a more intuitive interface for mouse behavior and for symbols for basic functionality. (cf. Iconicode). Why distinguish between rcv and vrcv, snd and vsnd? Why not have the system figure out the best way to send the data? ------------------------------------------------------------------------------- From: keeper@oregon.cray.com (Brian Whitney ext 2682) >Are you using PVM today? We are in the process of bringing it up. So far it has run on several SPARCs. >What are your application(s)? We manufacture computers and are investigating what PVM is and what kind of things it can do for us. >How many and what kind of machines have you used with PVM? Currently, we are running on SUN 470, SUN 670, SUN SS 1+, CRAY S-MP and CRAY APP machines. Brian Whitney Cray Research Superservers, Inc. (503) 641-3151 x2682 3601 SW Murray Blvd keeper@cray.com Beaverton, OR 97005 ------------------------------------------------------------------------------- From: "Ernest Retzel (1535 49118)" > Are you using PVM today? we have installed and are using it, tho it is not a production application in our environment. > What are your application(s)? dna and protein similarity searching. > How many and what kind of machines have you used with PVM? 6 Suns [not all we have, but the highest-performance, largest memory machines--SS2 and IPX] Ernie Retzel ernest@lenti.med.umn.edu ------------------------------------------------------------------------------- From: Friedbert Huber I'm just back from holiday, so I haven't used PVM very extensively. To your questions: 1. With high probability I will not use PVM today (=monday) 2. applications: writmng some simple PRAM algorithms to check whether PVM can be used for practical student exercises corresponding to a lesson on (theory of) parallel and distributed algorithms in winter semester (oct. 92 - feb. 93). 3. in the actual (test) phase I'll use 8 SUN sparc stations. I hope this helps you. Best wishes, Friedbert ------------------------------------------------------------------------------- From: parashar@npac.syr.edu > > Are you using PVM today? Yes ! > > What are your application(s)? Linear Algebra (LU factorization), 2D FFT, Quantum physics (current) > > How many and what kind of machines have you used with PVM? Machine Types: Sun's (Sparc 1+, ELC, IPC) IBM RS6000 DEC Workstations Max number: 8 =============================================================================== Manish Parashar Phone: 315-443-2483 (Off.) NPAC, 1-019 CST 315-426-0284 (Res.) Syracuse University parashar@nova.npac.syr.edu Syracuse, NY 13244. parashar@cat.syr.EDU =============================================================================== ------------------------------------------------------------------------------- From: mcs@memws.ssd.ornl.gov (Michael Stricklin) As a summer intern, my use of pvm has been limited to trying to educate the physicists I work with in the use of their parallel processing resources. I have written three sample programs, two in C and one in FORTRAN, and have run them on RS6000's and SUN4 SparcStations. No real research has resulted as of yet. Thanks for your interest. Michael C. Stricklin mcs@memws.ssd.ornl.gov strickli@cs.utk.edu ------------------------------------------------------------------------------- From: tmitsoli@shearson.com (Thanasis Mitsolides) > Are you using PVM today? No. But I would like to if it becomes less sensitive to workstation failures. > What are your application(s)? Financial, Servers, DBs, Controlers > How many and what kind of machines have you used with PVM? SUNs and HPs > Any comments and suggestions for improvement are welcome. > Please send the information to pvm@msr.epm.ornl.gov. > Your help is much appreciated. I really liked the simplicity, power and compatibility of PVM. The fault tolerance problem with the daemons though is serious. I have to say that it is impossible for us to use a system that terminates all programs using it whenever one of the daemons goes down. All the best to you. Thanasis ------------------------------------------------------------------------------- From: olson@juliet.ll.mit.edu ( Steve Olson ) > >Are you using PVM today? I'm doing some light experimentation. I was, in fact, using it the day the survey arrived. >What are your application(s)? Numerical weather analysis and prediction. >How many and what kind of machines have you used with PVM? Only Suns so far. We have some SGI's, I'll eventually try it out on those machines also. >Any comments and suggestions for improvement are welcome. >Please send the information to pvm@msr.epm.ornl.gov. >Your help is much appreciated. I have been fooling around a little bit with PVM and HeNCE. I have not used the system enough to give any detailed comments. Both PVM and HeNCE compiled easily, but both seem awfully picky about how installation directories have to be arranged. My #1 priority is to have a package that is easy to use, particulary with pre-existing applications. I'm not doing parallel processing research, I'm trying to make a existing application run fast on a network of Suns and SGI's. -- Steve Olson -- MIT Lincoln Laboratory -- olson@ll.mit.edu -- ------------------------------------------------------------------------------- From: lhdsy1!mailmaster.chevron.com!ussr01.chvpkh.chevron.com!hdgki@uunet.UU.NET (Dave G. Kissinger) I am just starting to look into PVM, and hope to use it for "batch" processing of seismic data on our 12 DEC 5000 workstations. ,. ------------------------------------------------------------------------------- From: Vincenzo De Florio > Are you using PVM today? Yes. > What are your application(s)? I developed a simple numerical integration routine, and I'm trying to port a parallel edge detector from a MEIKO Computing Surface environment to PVM. > How many and what kind of machines have you used with PVM? Thanks to PVM I integrated four machines i.e., two RIOS, a SUN3 and a HP9K. A Computing Surface, attached to the SUN3, offers us eight T800 transputer and a frame grabber that we're trying to integrate into a single PVM environment. > Any comments and suggestions for improvement are welcome. I've already sent a simple suggestion concerning the need for a dynamic load balancing feature in PVM. If you want that an applications take real advantage from a pool of computational elements in a PVM environment, it must know how the computational power is distributed. Best wishes, Vincenzo De Florio Tecnopolis CSATA Novus Ortus -- Italy ------------------------------------------------------------------------------- From: genda@lily.ics.agh.edu.pl (Andrzej Krol) Yes I am and my colleague are using PVM today. We use PVM to parallel computing on set of workstations : - ray tracing, fractals, test loading net; transmition speed, molecular dynamics (compare efficiency with other solutions eg. using STRAND, ANSA) I use with PVM SUN Sparc stations (SLC, IPC, SPARCstation 2, SPARCserver 470), and IBM RS 6000 (320). If you have any more questions, ask me. Best regards, Andrzej Krol | genda@lily.ics.agh.edu.pl University of Mining and Metallurgy Institute of Computer Science 30-059 KRAKOW POLAND ------------------------------------------------------------------------------- From: "Joefon Jann" Are you using PVM today? Yes, I am trying to learn how to use it. What are your application(s)? Various mathematical problems e.g. matrix diagonalization, cellular automata, etc. How many and what kind of machines have you used with PVM? RS/6000 for now. Will PVM run on AIX/ESA on IBM's ES9000 eventually? ------------------------------------------------------------------------------- From: mikem@stat.cmu.edu > Are you using PVM today? We are doing some small experiments, but no production work. > > What are your application(s)? Simple ones, we will gear up for some more complicated statistical problems, eventually. > > How many and what kind of machines have you used with PVM? We have a network of about 30 decstations, and those are the machines we are using. The communication is over ethernet. At some stage we may move to a higher-speed network, perhaps FDDI, but that is still a while off. Regards, --Mike Mike Meyer, Computing Services and Department of Statistics, Carnegie Mellon University ------------------------------------------------------------------------------- From: kpc23@cas.org (Kevin P. Cross Ext. 3192 Room 2209B) Are you using PVM today? yes with Hence - the only way to use it! What are your application(s)? To demonstrate the distributed computing environment, its advantages and problems. Chemical cluster analysis and database searching are two application areas. How many and what kind of machines have you used with PVM? So far, 10 machines, SUN SS2, 4/690, SS1. However, we have approx. 1000 total UNIX machines: (most)SUN, IBM, DEC, HP, SGI. ================================================ Kevin P. Cross, Ph.D. Research Chemical Abstracts Service 2540 Olentangy River Road P.O. Box 3012 Columbus, OH 43210 kpc23@cas.org Internet kpc23@cas Bitnet (800)-848-6538 Ext. 3192 ================================================ ------------------------------------------------------------------------------- From: wrj9493@lippy.boeing.com (Wayne Jones) I am responding to your request to Jeff Lewis (jpl8832@rex.boeing.com) for info on our PVM usage. We are in the same group and I am the focal point for our PVM activites. We have just begun trying out PVM recently and the target application is a CFD code utilizing modular multizonal methods which lend themselves well to coarse grained parallelization. Initially, we are running it on a cluster of 12 SGI "user owned" workstations. Later I expect we will extend it to include some RS6000's, some HP 700 series machines and our Cray YMP. In the nature of comments and suggestions, I offer the following: 1) Of general interest to us are procedures to utilize spare cpu cycles on our workstations. We were in Tallahassee last week and spoke with Tom Green and Jeff Snyder about their DQS work and I expect we will try that out. However, their approach to supporting parallel work using PVM by allowing requests for multiple cpus is not quite what I had hoped for with our application. Some sort of dynamic allocation of cpu resources would be more appropriate. The reason is that in our application, while a maximum number of cpus needed can be predetermined, it is rare that all would be needed at the same time. The number needed depends on the state of the solution and I would prefer not to tie up cpus not in use. For our initial efforts I just trivially cycle through a list of available machines but there is no "search for least loaded machine", accounting or anything else. So, question one for you is -- do you know of any more elaborate appoaches to dynamic allocation of cpus for PVM usage? I made some rudimentary efforts a couple of years ago to deal directly with socket calls to handle some interprocess communications. After that experience, it is a great pleasure to work with PVM instead. Keep up the good work. Wayne Jones ------------------------------------------------------------------------------- From: cheekong@iss.nus.sg (chui chee kong) Thank you for delivering pvm to us. I think that this is a good piece of software. It has helped us in our research work. These are the feedback we had for PVM. Are you using PVM today? What are your applications ? Yes. A group of 3 undergrads are using it to implement distributed ray casting. We will also be using it to solve large scale optimization. My collegues are going to use PVM for medical imaging and volume rendering in the near future. How many and what kind of machines have you used with PVM? 19 Sun's, some Sparc2's, some IPC's, some Sparc1+'s. Suggestions: * should have a debugging environment as currently tracing the variables using printf is tedious. * will be good to set the prority of execution of tasks in the remote stations. Problems: * seem like pvm does not permit the sending of a very large block of data (~ 2 MB) and most of the time the program will have error vxfer.... or message is dropped. Currently, avoiding this problem by partitioning the data into bits to send. * cannot get pvm to run across a network of Sun stations running SUNOS release 4.1 and Silicon Graphics IRIX workstations. If you have any progress on PVM, please let us know. Thank. chee kong :-) internet: cheekong@iss.nus.sg ------------------------------------------------------------------------------- From: Morten Bjoerhus I am a PhD student of Professor Syvert P. Norsett at the Norwegian Institute of Technology. I am working on an iterative technique for first order systems of hyperbolic PDEs, making use of domain decomposition. Just a month ago I made the first parallel version of my program, using the PVM software. I was very surprised of how easy it was to understand and use the PVM software, and in fact I didn't use more than three days to make a parallel version of my sequential C code. In addition, my algorithm is very coarse grained, i.e. communication is negligible as compared to the computation done on the different nodes, so PVM seems very suitable for my purpose. The reason I am writing to you is really because Gene Golub urged me to, as he meant that you would like to hear some experiences from a PVM user. He is here in Trondheim this week attending the conference "Industrial Mathematics Program: The first ten years - experiences and challenges", which is organized by my department, the division of mathematical sciences. In fact, I also had the pleasure of being given "The Gene Golub prize" for the best talk given by a PhD student at the conference; which was a talk about my rather preliminary research on this technique, but in which I also referred to my numerical experiments with PVM on the Suns at our department. So I really hope this gave the audience a good impression of your PVM package, as I think it deserves. So this was just an attempt to give you some positive feedback on the PVM software, and I hope you will put me on your PVM mailing list (if such a list exists) so that I can benefit from new releases or developments. Best regards, Morten Bjoerhus. ------------------------------------------------------------------------------- From: souva@aibn55.mpifr-bonn.mpg.de (Ignatios Souvatzis) Using today: yes. Applications: solving lots of similar ode's in parallel. Machines: 6 DECstations, 1 Sparcstation, 2 Convex C1's Suggestion: I woulld like a windowed version of the interactive PVMD interface. No, not like Hence; Hence places too many restrictions on the data flow graph for me. I don't want to restart a Unix process for each ODE I want to solve, only for each machine I use to solve them. A Hence version with less dataflow restrictions would be nice, too. ------------------------------------------------------------------------------- From: Xavier-Francois.Vigouroux@lip.ens-lyon.fr (Xavier-Francois Vigouroux) 1. I am not using PVM today because I'm on hollydays. 2. I use PVM to simulate a PM that doesn't already exist. The program is a monitoring program. 3. Only one kind of machines : SUN-4 (SPARC). ------------------------------------------------------------------------------- From: weber@zam004.zam.kfa-juelich.de (Michael Weber) 1. Yes I'm still using PVM. 2. I'm in a state of looking for a real application suited to run with PVM. It's not so easy to find users who are to convince to rewrite their serial code to a parallel one (Fortran of course) if you can't promise them an tremendous speedup. Up to now I only used your example programs and some simple test progams we wrote to test communication. 3. The core of my installation are four IBM RS/6000 modell 320 H. They are connected via ethernet. This installation works stable and after some NFS mounts is easy to handle. For reasons of testing I also integrated a DEC and a SUN workstation and a CRAY XMP into the PVM - this did work too. This point is interesting one for me. As we also are equipped whith an iPSC/860 it would by interesting to integrate it in an application together with the CRAY and some workstations. 4. What's about a more actual users guide? Michael Weber Forschungszentrum Juelich GmbH ZAM Postfach 1913 5170 Juelich Germany Tel.: 02461-612339 Fax.: 02461-616656 email: weber@zam004.zam.kfa-juelich.de ------------------------------------------------------------------------------- From: frese@sc.ZIB-Berlin.DE (Hans-Hermann Frese) Thank for your questionnaire on PVM usage. Below I have tried to summarize the use of PVM at our institution. > Are you using PVM today? Yes, we are using it since end of 1991. > What are your application(s)? (1) We used PVM for development of message passing programs, mainly for benchmarking distributed memory MIMD machines. (2) We shall use PVM in the development of a framework for combinatorial optimization (branch & cut algorithm). (3) We shall use it for production runs on workstation clusters (Sun, HP, SGI, may be in concert with a Cray Y-MP) for large problems of combinatorial optimization. (4) We are also investating the use of PVM for distributed supercomputer applications. > How many and what kind of machines have you used with PVM? We used it on up to 9 machines concurrently (Sun-3, Sun-4 and SGI-Machines). Some of our benchmark programs run on MIMD machines with PVM implemented on top of their own message passing library, employing up to 32 processors. Distributed supercomputer applications will run on 2 CRAYs and a front-end workstation (Sun SPARC or SGI). > Any comments and suggestions for improvement are welcome. Have you thought about including interrupt driven communication a la intel in PVM? We find it a usefull extension to standard send/receive calls. Please keep us informed about the future development of PVM. Thank you. Regards, Hans-Hermann Frese Konrad-Zuse-Zentrum fuer Informationstechnik Berlin Heilbronner Strasse 10 D-1000 Berlin 31 Germany Phone: +49 30 89604 145 Fax: +49 30 89604 125 Email: frese@ZIB-Berlin.de ------------------------------------------------------------------------------- From: refloch%FRONER81.BITNET@pucc.Princeton.EDU I am Alain refloch, i am the chief of the user's support team we have in ONERA ( office national d' etude et de recherche aerospatiale ) ( electromagnetic field and a big part of fluids mechanics ) a Cray Y-MP 4-128, and a Intel iPSC 860 (128 nodes) Yes i am using PVM today, for small application, and for tests I am using PVM on the Cray ( only on the Cray ) and on a Cray and a Sun ( remark : for big application, which run in batch, it is necessary to use the Korn shell on the cray, because with the bourne shell the nice value of the pvm daemon is 5 less of the processes fhost-nodes and the job rest a very long time in machine if there is a lot of jobs in run I coach a presentation in october about pvm, for the users. I think that pvm is very interesting, and for me one of the principal reason is the portability, between a network of Sun, a Cray , or a iPSC, a other reason for the future, is that pvm is supported by the programming model of the computer MPP of Cray in 1993. For users, it is not easy to write big applications, for Cray and after for iPSC, because the user waste one's time to translate the program , I think that pvm is the ideal tool for the users to begin the migration between the vector supercomputer, and the parallel computer. I am sure that pvm have a great future. bye refloch@froner81.bitnet ------------------------------------------------------------------------------- From: LT3NAIM2%CINE88.CINECA.IT@UTKVM1.UTK.EDU We are using PVM release 2.3 from three months for CFD application developped on massively parallel computer (i.e. INTEL , transputer network) . Our usage of PVM is performed on a cluster of 6/8 IBM R6000. We hope to have in new releaser of PVM a complete set of non blocking trasmissions. Best Regards PASQUALE SCHIANO Italian Center for Aerospace Research Via Maiorise CAPUA , ITALY ------------------------------------------------------------------------------- From: Luigi Brochard We are an IBM team in France which works helping our customers parallelizing their applications on network of RS/6000 connected with Ethernet and SOCC (with a version of PVM adapted to SOCC done by IBM ECSEC in Rome). The applications we have worked on today are from CFD and seismic. In the area of improvments: . a higher level of communication library for handling matrices . an X interface . a debugger Thanks for PVM, Luigi ------------------------------------------------------------------------------- From: roar@sima.sintef.no (Roar Skaalin) > >Are you using PVM today? > Yes! It's used in research, but it will also be used in cources on parallell computing at the Norwegian Institute of Technology. >What are your application(s)? > So far, in-house applications in marin hydrodynamics and research on algorithms for molecular dynamics >How many and what kind of machines have you used with PVM? > IBM RS600, DECstations (up to 20 at a time), SGIs and CRAY Y-MP. More systems will be used later. Among them 40 SUNs. Have you tested PVM on the Intel Paragon yet? I wish you good luck in your further work on PVM. *************************************************************************** * Roar Skaalin Phone: +47-7-59 30 45 * * SINTEF Industrial Mathematics Fax : +47-7-59 29 71 * * N-7034 Trondheim email: roar@sima.sintef.no * * NORWAY * *************************************************************************** ------------------------------------------------------------------------------- From: Sharon Smith I am currently using PVM for several applications. The first application involves a project that will couple a model of the ocean and a model of the atmosphere. PVM is being used to send the boundary information between the two models. This work is part of a major effort in France to couple these two models and Norman Barth is the main scientist involved in this work. After reading your request, I sent a technical report to Jack Dongarra to give more information about this. The second application is a parallel algorithm for solving the global optimization problem. This algorithm is very coarse grained and extremely well suited for a heterogeneous computation environment. We are using the algorithm to solve optimization problems for planning satellite trajectories for satellites dedicated to collecting altimeter information. The first application is currently being developed for the Cray 2. The second application runs on a network of SUN workstations, although we have plans to incorporate other workstations, such as the IBM RS6000 and the SGI machine. I hope that this information is helpful. I am happy to answer any further questions that you might have concerning these efforts. Sharon L. Smith Cerfacs phone: (33) 61 19 30 23 42, Avenue Gustave Coriolis e-mail: smith@orion.cerfacs.fr 31057 Toulouse Cedex France ------------------------------------------------------------------------------- From: NE12%DKAUNI2.BITNET@vm.gmd.de Here are the answers to your three questions. I have a few questions with the hope it will be not too much work to answer my questions by e-mail (ne12@rz.uni-karlsruhe.de). Are you ...? Yes (not always, but always more often) What are ...? Numerical algorithms (e.g. Gauss-algorithm) How many ...? From 3 up to 16 HP-workstations (model 720) Suggestions: - the possibility to separate easily the outputs of the different nodes and/or - the possibility to equip the outputs with a time-stamp (perhaps the outputs could be sorted by the time) - explanations to the messages which the PVM-daemon writes to stdout (e.g. ...getndfloat: END OF BUFFER ) Questions: - Can you send me explanations to the above mentioned messages - Do you have a paper about the internal organization of PVM. Especially I would like to know how the buffers of the PVM-daemon are managed and how 'broadcasting' works (due to resending of messages). Best wishes (chiao,tschuess) Hartmut Haefner ------------------------------------------------------------------------------- From: Richard Bielak > > Are you using PVM today? Yes. > > What are your application(s)? > I'm trying to use PVM to write a concurrency library for Eiffel (an OOPL). PVM is easy to use as for transport of objects between processes, and for asynchronous invocation of methods. > How many and what kind of machines have you used with PVM? We're using RISC/6000s and DECstations. > > Any comments and suggestions for improvement are welcome. More docs and more examples :-). Some tools for tracing and timing the flow of messages. For example, it would be nice to be able to dump all messages send and received by a program into a file. A dynamic display of the messages would also be very nice. -- * Richie Bielak (212)-815-3072 | "Your brain is a liquid-cooled parallel * * Internet: richieb@bony.com | super-computer". He pointed to his nose, * * Bang {uupsi,uunet}!bony1!richieb | "This is the fan." * * - Strictly my opinions - | - David Chudnovsky - * ------------------------------------------------------------------------------- From: eltgroth@ocfkms16.llnl.gov (Peter G. Eltgroth) I am running a very large atmospheric circulation code on an IBM cluster using PVM2.4.1. All seems to be fine, and I thank you for your good work. Pete Eltgroth LLNL, L-298 Livermore, CA 94550 eltgroth@physics.llnl.gov ------------------------------------------------------------------------------- From: "Dr. C.D. Wright" We are not using PVM today, although work is continuing on building an inplementation of PARMACS on top of it. We have no applications built on PVM yet. We have used 10 SUN workstations on an ethernet with PVM. Regards, Colin Wright. ------------------------------------------------------------------------------- From: "Christian P. Roberts" > > Are you using PVM today? Yes, Version 2.4.1, and am trying to get Xab. > > What are your application(s)? Converting a hypercube code that compares a protein sequence to a database of protein sequences. > > How many and what kind of machines have you used with PVM? I have compiled it and used the samples between and among RS/6000 model 550's and Sun SPARCStations. > > Any comments and suggestions for improvement are welcome. I think I communicated this to you earlier - that when I was in hopes of establishing a link to the installation at the Pittsburgh Supercomputing Center, I was unable to do so - because the installation they have is version 2.4.0, and that can't talk to version 2.4.1. We see this as a major problem, since folks can't hope to have the same version on their own machines and ones they might want to use over the network, outside their own camps. Chris Roberts Academic Computing Center cpr4k@virginia.edu University of Virginia (804) 982-4693 Charlottesville, Virginia 22903 ------------------------------------------------------------------------------- From: telaydi@george.lbl.gov (Tarek Elaydi [ICSD Summer Program]) Are you using PVM today? A: Yes, it is now the center of my project. What is your application? A: it is a raytracer using constructive solid geometry to model scenes. How many and what kind of machines have you used with PVM? A: I usually run the application on 6-8 machines. Most of the hosts are Sun Sparc II's, working with two Sun-690MP's and one SGI-Crimson. As soon as I polish up the application, I will try adding a CM-5, Cray Y-MP and if possible an Ardent st3000. Any comments and suggestions for improvement are welcome. Currently, I manually initiate components using information from a dynamic load balancer I wrote. It has been suggested to me that such a feature would be helpful if it was part of PVM. In other words, when one would call initiate, PVM will call each host for a perfromance stat structure and use that info to decide where the next component goes. I thought that an older version of PVM did this by using load as a metric. How does initiate decide where the next component goes now? I am willing to work closer with you on this if need be so e-mail on what you think. Thanks Tarek Elaydi telaydi@george.LBL.gov ------------------------------------------------------------------------------- From: pm@fct.unl.pt (Pedro Medeiros) >> Are you using PVM today? Yes. >> What are your application(s)? Our group is using PVM as one of the basis for the integration of distributed-memory multiprocessors in local area networks. We envisage applications in the areas of simulation of industrial processes and visualization. >> How many and what kind of machines have you used with PVM? About 15 workstations (DEC Vaxstations and Decstations, IBM RS/6000, HP/720, Sun Sparc ). We are porting PVM to a Meiko Computing Surface ( a Transputer-based multiprocessor). Our system is hosted by a SparcStation, has 16 CPUs and is running the Meiko CSN communication system. E-mail: pm@fct.unl.pt FAX: + 351 1 295 56 41 S-mail: Pedro Medeiros Dept. of Computer Science, Universidade Nova de Lisboa 2825 Monte Caparica, Portugal ------------------------------------------------------------------------------- From: rwb@cvx.crd.ge.com (Robert W. Benway) I now *seem* to have PVM running on the HP 750. It was by *far* the worst machine I've put PVM on to date (including: CRAY-2, CRAY-XMP, Convex [240 & 3220], SGI personal Iris, SGI Indigo, various DECstations, and an Intel Hypercube). We are beginning to see the real *potential* of the beasty. My sincere congratulations on your group for having the foresight to come up with this entity. (Actually, I'm just a bit envious....wish I had thought of it...) BTW, I reworked all the f2c routines to add the "ifdef HP9K" steps ala RIOS. (The area where we are currently using PVM is in working with finite element problems in the electromagnetics area.) Most of our work (99.99999%) is in fortran. Hence, the interest in f2c.... BTW2, is there a schedule for PVM 3.0 ? and will it run on a Hypercube, taking advantage of the cube itself (nodes)??? (Any info would be helpful.) Thanx very much again, -RWB ------------------------------------------------------------------------------- From: GERADIN / Albert Lerusse u5174a1@bliulg11 / / Universite de Liege / / Laboratoire de Techniques Aeronautiques et Spatiales / / Service de Dynamique des Constructions Mecaniques / / 21, Rue Ernest Solvay / / B-4000 LIEGE / / BELGIQUE / / tel 32-(0)41- 66 91 58 / / Fax 32-(0)41- 52 10 49 / ======================================= PVM is installed today on my network of workstations and I have realized tests with PVM. ======================================= My applications are finite elements application, I developpe a method to study the dry friction phenomena by a multi harmonics method. This work is supported by European Communities. Like I begin to work on parallel computer, I have first realized short tests. I send you a postscript copy (include in this file) of my first report. This report describe the tests realized with PVM. ======================================= My network of workstations is composed by 2 Hp 9000/730 + 1 Silicon Graphics Indigo + 1 Silicon Graphics 4D-power (4 cpu) in 1 month this network would increase with 3 Hp 9000/710. The only problem with this network is that we don't have optical fiber network, we use a simple ethernet cable. ======================================= Suggestions: -1) a debugger for PVM -2) a list of persons working with PVM and the research subject to obtain interaction between users. I would thanks all peoples who work to realise PVM. A. Lerusse ======================================= postscript file [could not be printed-Al] ======================================= ------------------------------------------------------------------------------- From: bking@earth.cray.com (William King) Are you using PVM today? Only for simple prototype applications What are your application(s)? Simple examples of distributed applications which perform basic signal processing functions. How many and what kind of machines have you used with PVM? Multiple Cray platforms and Sun's. Bill King Cray Research, Inc. 612-683-7230 bking@cray.com ------------------------------------------------------------------------------- From: SDT Advanced Development Manager DTN 381-2084 PVM seems to mostly be used for converting images from one format to another (including things like color quant, scaling, etc.) I rarely use it for image cleanup, etc. because I prefer an interactive tool for that. I use PVM only on Ultrix (DECstations mostly). I would like to see more formats supported, particularly .bmp and .jpg A GUI shell would be nice--the pipe thing is a little tedious once you get used to things like the tools on windows3. The Khoros folks (and the earlier Xvision) were able to use pipes as the underlying structure and then place a GUI on top. It would also be nice if you incorporate our DECs ".ddif" converters (ddif2ppm) into the public release. ------------------------------------------------------------------------------- From: David M woodcock I put it on sun sparc,dec5000,rs6000,hp 9000s [all afs] here at CAEN. Other than a few test programs and additional layer [rhapsody] havent been building pvm prrograms myself tho several grad students here have. suggest you email idaho@engin.umich.edu do you havea ksr vversion? ------------------------------------------------------------------------------ From: jmw@random.asd.sgi.com (Jim Winget) Sorry for the short reply, but I'm in the middle of "Hot Chips". I've "ported" several versions of pvm since its public release two (?) years back. All of my work has concentrated on the Silicon Graphics IRIX multiprocessor environment. Unfortunately, I have not yet had time to implement the needed "second" level of machine support to fully utilize the shared memory available when running mutiple copies of pvm clients on an MP machine. The obvious question is whether to expose this additional structure through the user interface or hide to provide a more homogeneous environement. Has the project team made any advances in this direction? Is there an "internals" guide? Thanks, Jim --- James M. Winget Voice: (415) 390-3654 Silicon Graphics, Inc. Email: jmw@sgi.com Research & Development, 7U-005 FAX: (415) 965-7651 2011 N. Shoreline Blvd Mountain View, CA 94039-7311 ------------------------------------------------------------------------------ From: pereira@mbeya.research.att.com (Fernando Pereira) 1. I've not started using PVM yet. 2. The intended application, which I expect to try in the next few months, is training of statistical grammars from examples. The algorithms are already implemented on Sun SPARC and Alliant FX/2800, but I plan to build new distributed versions using HeNCE. 3. Not yet started, but the intended machines are Sun SPARC, Alliant FX/2800 and SGI Indigo. Fernando Pereira 2D-447, AT&T Bell Laboratories 600 Mountain Ave, PO Box 636 Murray Hill, NJ 07974-0636 pereira@research.att.com ------------------------------------------------------------------------------ From: ramus@nersc.gov (Joe Ramus) I am not using PVM. But other people at our site are using it. Joe Ramus NERSC Livermore (510) 423-8917 ramus@nersc.gov ------------------------------------------------------------------------------- From: alex@irs.phy.nrc.ca (Alex Bielajew M-35 X-12) > Are you using PVM today? No, we are just getting started. > What are your application(s)? We are EGS (Electron Gamma Shower) developers with a user base of about 5000. This community spans High Energy Physics, Radiation Protection, Radiation Research in general and Medical Physics. EGS is a public-domain Monte Carlo code that does electron gamma transport in the energy range 1 keV -- 20 TeV. It is ideally suited to a network operating system like PVM since it requires almost no internode communication. > How many and what kind of machines have you used with PVM? We have 7 Sparc II's under our direct control, free use of an 8 CPU SGI compute server. When we get PVM going we will be able to reign in another 200 or so machines of this class that are on the network of our organization. All workstation-class vendors are represented at our site. > Any comments and suggestions for improvement are welcome. > Please send the information to pvm@msr.epm.ornl.gov. When we have intelligent questions we will contact you. > Your help is much appreciated. Thank YOU for your wonderful and timely contribution to computing science. If you have anyone there running EGS on PVM, please initiate some communication. Gratefully, Alex F Bielajew Institute for National Measurement Standards National Research Council of Canada Ottawa, CANADA K1A 0R6 tel: 613 993 2197 fax: 613 952 9865 e-mail: alex@irs.phy.nrc.ca ------------------------------------------------------------------------------- From: mk@prosun.first.gmd.de (Matthias Kessler) >Are you using PVM today? Yes. >What are your application(s)? We are working on a parallelizing FORTRAN Compiler and generate C Programs which run on top of PVM. We use some numerical applications and let them run under PVM to test our Compiler. At this time we only test the correctness of the generated programs, we don't care about the speed of PVM. >How many and what kind of machines have you used with PVM? We have a net of SUN4 workstations. Something like 100 machines, but we only use 10 at a time for PVM applications. So far this setup has worked very well. Thanks for a great tool and kepp on the good work. At the moment we don't have any new suggestions, we're very pleased with what we have now! Matthias