IBM System x3755 M3



No CUDA, NAMD multicore 2.9

1,632 atom system

Cores Nanoseconds per day
1 22
4 69
8 99
12 113
24 131
32 117
48 36



6,342 atom system

Cores Nanoseconds per day
4 14
12 37
24 51
36 53
48 14



24,982 atom system

Cores Nanoseconds per day
12 9.3
24 15.4
32 14.3
36 15.0
48 16.9
48, PMEProcessors=24 17.8
48, PMEProcessors=8 17.3



60,626 atom system

Cores Nanoseconds per day
12 3.24
24 5.7
36 5.6
48 6.9



ApoA1 system

Cores Nanoseconds per day
48 2.25






With CUDA, NAMD multicore_cuda 2.9

The card (quadro K2000D) is -unfortunately- not up to required standards, but it was possibly the only reasonable solution : not only the space in the riser PCI assembly of the server is very tight (requiring a slim card), but the 16x slot can only take up to 70 Watts worth of a card. This means GTX 645 (at ~120W) wouldn't do it, and K4000 (rated at 80W) was too expensive at ~$800.


Anyway, here are the numbers :


6,500 atoms

04 cores : without CUDA → 15.2 ns/day , with CUDA → 12 ns/day


25,000 atoms :

04 cores : without CUDA → 3.64 ns/day , with CUDA → 6.90 ns/day

08 cores : without CUDA → 6.96 ns/day , with CUDA → 7.26 ns/day

16 cores : without CUDA → 11.62 ns/day , with CUDA → 7.27 ns/day


60,000 atoms :

04 cores : without CUDA → 1.29 ns/day , with CUDA → 2.70 ns/day

08 cores : without CUDA → 2.43 ns/day , with CUDA → 2.97 ns/day

16 cores : without CUDA → 4.50 ns/day , with CUDA → 3.10 ns/day



The way it looks, it would possibly make more sense to place the card in an old quad-core box (n0006 ?) where it would come handy for a simulation with a large number of atoms, instead of wasting it on the IBM server (why running a job with only 4 or 8 cores in a 48-core machine ?).

⇒ Take-home message : stick to simple boxes (with ~16 cores) with very fat nvidia cards. For current versions of NAMD they can produce as many ns/day as a 48-core machine without a fat nvidia card.



about/benchmarks/ibmx3755m3.txt · Last modified: 2013/09/11 20:46 by glykos