On Wednesday 09 April 2008 11:48:19 am Greg Lindahl wrote: > p.s. did anyone see this blog posting claiming that it takes Linux > clusters several minutes to start a 2048 core job? > > http://terboven.spaces.live.com/Blog/cns!EA3D3C756483FECB!267.entry > > I would have commented, but you have to have a Microsoft ID to > comment.
That's a pretty amusing blog, actually. See http://terboven.spaces.live.com/blog/cns!EA3D3C756483FECB!255.entry (wonderful permalinks, btw). The same benchmark on lower-end hardware (E5345) running Linux (same 4X DDR IB though), gives roughly 30% better results: #--------------------------------------------------- # Benchmarking PingPong # #processes = 2 #--------------------------------------------------- #bytes #repetitions t[usec] Mbytes/sec 0 1000 4.72 0.00 1 1000 4.68 0.20 2 1000 4.64 0.41 4 1000 4.74 0.81 8 1000 4.68 1.63 16 1000 4.78 3.19 32 1000 4.79 6.37 64 1000 4.84 12.61 128 1000 6.12 19.95 256 1000 6.47 37.71 512 1000 7.14 68.40 1024 1000 8.53 114.54 2048 1000 10.81 180.65 4096 1000 14.12 276.58 8192 1000 20.99 372.21 Graph at: http://tinyurl.com/4ndg9v Note that latency scaling is much more linear, too. So much for Windows better performance and lower latency, huh? Cheers, -- Kilian _______________________________________________ Beowulf mailing list, Beowulf@beowulf.org To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
