Hallo Jan,
Samstag, 19. April 2008, meintest Du:
|
> |
KC> That's a pretty amusing blog, actually. See KC> http://terboven.spaces.live.com/blog/cns!EA3D3C756483FECB!255.entry KC> (wonderful permalinks, btw). Read this arcticle to the end: "Finally, lets take a quick look at some performance measurements. From the Intel MPI Benchmark Suite V3.1 I used the MPI-1 part to measure the latency and bandwidth. Please note that you have to divide the PingPong-latency by two in order to get the "real" MPI-latency. In doing so we see a MPI latency of aboud 2.9 us on our 4x DDR InfiniBand. " What? They divided the number of the benchmark by two? Correct me if i'm wrong but i was always assuming that the benchmark does this himself! Can somebody confirm that? |
So i answer myself:
ftp://ftp.uybhm.itu.edu.tr/belgeler/sistem_kullanimi/IMB_ug-3.0.pdf
Page 10 Chapter 3.3.2.1 PingPong: "time = Δt/2 (in μsec) as indicated in Figure 1". So the tool already shows just the one-way-time. No division necessary.
Cheers,
Jan
_______________________________________________ Beowulf mailing list, Beowulf@beowulf.org To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
