PSTSWM AlphaSC-667 Serial Performance

Performance Studies using

PSTSWM

Date/Person:	April 17, 2000 / P. Worley
Platform:	Compaq AlphaServer SC at Oar Ridge National Laboratory (colt.ccs.ornl.gov):
	16 ES40 4-way SMP nodes (667 MHz Alpha 21264a with 8MB L2 cache)
Environment:	Digital UNIX V5.0; RMS 2.36
Code Version:	7.2
Make Options:	MACH=alpha-sc COMM=mpi PRECISION=8 PERF=n WORKSPACE=20000000 MATH=cxml
Compilation Options:	f90 -O4 -assume accuracy_sensitive -math_library accurate -align dcommons -align records -arch host -tune host
Number of steps:	T5, T10, T21, T42: 241 or 481
	T85: 49 or 97
	T170: 49 or 97
Number of processors per node:	1, 2, 3, or 4
Notes:	using CXML library routines for Fourier transforms and BLAS

T5 MFlop rates

Per Processor Rate

Per Node Rate

**T5 MFlop rates**
Per Processor Rate	Per Node Rate

T10 MFlop rates

Per Processor Rate

Per Node Rate

**T10 MFlop rates**
Per Processor Rate	Per Node Rate

T21 MFlop rates

Per Processor Rate

Per Node Rate

**T21 MFlop rates**
Per Processor Rate	Per Node Rate

T42 MFlop rates

Per Processor Rate

Per Node Rate

**T42 MFlop rates**
Per Processor Rate	Per Node Rate

T85 MFlop rates

Per Processor Rate

Per Node Rate

**T85 MFlop rates**
Per Processor Rate	Per Node Rate

T170 MFlop rates

Per Processor Rate

Per Node Rate

**T170 MFlop rates**
Per Processor Rate	Per Node Rate

The performance discrepancy here indicates the cost of contention for memory (bandwidth) when multiple processors are running simultaneously. The total memory required per node goes up as the number of processors increases, but the per processor cache behavior and memory access patterns is constant.
This experiment indicates the difference in performance one might see in a space sharing context, sharing the node with other users, or in a scaled speed-up experiment, where the size of a parallel job is increased as the number of processors used increases.
Note that two different metrics are displayed - per processor computational rate and per node computational rate. The per node rate is the best measure of the performance of the architecture, but the per processor rate is a convenient way to observe performance degradation.
Summary: Per processor performance degrades as the number of processors used per node increases for all of the larger problem sizes. The performance difference between 1 processor and 4 processors per node is fairly consistent for these cases, ranging from 50-100 MFlops/sec in most cases, with the 2 and 3 processor rates showing approximately one third and two thirds of the degradation of the 4 processor rate, respectively. Note, however, that it never makes sense to use less than 4 processors per node when looking at per node computational rates.