Last edited on 2011-06-06 21:26:57 by stolfi

Cache/Memory Performance Benchmarks of
Shared-Memory Multiprocessor Machines

Plots for alternative tests

Here are the results of other runs of the same benchmark, on slightly different machines:

dec.a8400c: DEC (Compaq) server, 8 CPUs alpha 21264 (EV6) 524 MHz, 4 MB on-board cache plus 64+64 KB on-chip cache.
sgi.flurry: SGI Origin, 124 CPU MIPS R12000 300 MHz, 8MB on-board cache, 32+32 KB on-chip cache (used in shared mode, with an unknown number of CPUs allocated by software to the benchmark).
sun.oslo: Sun SPARC Ultra-Enterprise HPC4500, 16 CPUs sparcv9 400Mhz, 4MB cache (configured as a 10 CPU system by software).

Actually DEC provided only one set of results, which was shown in the main page and repeated here for comparison.

Note that the results for the test sun.oslo are similar to that of sun.buddy on the main page, except for the reduced number of CPUs (10 rather than 20).

The results for sgi.flurry are much worse than those of the main test sgi.momentum. In particular the compact tests show no speedup for 2 to 10 threads. It is not clear whether these differences are due to the increased number of processors (124 instead of 32), which may have caused the active processors to be located farther from the process memory; or to the fact that this test was performed concurrently with other users.