PVM Performance Graphs


PVM communication bandwidth over a network (ethernet, FDDI, ATM, HiPPI) is equal to ttcp, which is the best one can obtain using TCP/IP supplied by the vendor. One way latency on these networks ranges from 1ms (ethernet, ATM) to 1.7ms (FDDI, HiPPI).

The results of the following graphs shows that the performance of pvm_psend is very close to the native MPP routines, while still maintaining portability and the ability to send messages off the MPP (to another MPP or a workstation).
Note that in PVM 3.3, pvm_psend only works between machines of the same data format. This restriction will be lifted in PVM 3.4.

The graphs also show that on MPP pvm_send is between 2X and 6X slower than pvm_psend. We are trying to understand and improve pvm_send on MPP, but for now pvm_psend is the clearly the routine of choice.

The following graphs are from a UT technical report by Henri Casanova and Jack Dongarra titled The Performance of PVM on Massively Parallel Processing Systems.
They used PVM 3.3.8 in their tests.