With 2 nodes, LAM will use a simple scheme -- the non-root will send to
the root, and the root will copy the sendbuf to the recvbuf.
I don't know offhand what would cause the results shown below, although
we have seen issues with the memcpy's in some OS's/Linux distros (i.e.,
some work better than others), particularly if the data it is copying
is page- or max-datatype-size-aligned or not (e.g., aligned on 8 byte
boundaries).
On Nov 11, 2005, at 2:49 PM, David Ulrich wrote:
> Hi,
>
> I try some basic benchmark on a small cluster. It's P4 2,4GHz with
> 83842 ethernet card (mtu 1500).
> With 2 nodes I have this:
> http://aspirine.li/mesure2.pdf
>
> on Ox it's the size of the packets in bytes, on Oy it's time in seconds
>
> We see very well the mtu effects, but I can't explain the second
> curve!?
> This is what I have done:
> - I use a float table which grow every loop
> - I send a table 100 times to do get a good time benchmark
>
> This is a part of the result, where we see that every 10bytes I have
> strange results:
>
> nodes;time[s];size [bytes]
> ------------------------------
> 2;0.000358077;16404<--
> 2;0.000343283;16408
> 2;0.000342555;16412
> 2;0.000342477;16416
> 2;0.000342885;16420
> 2;0.000342765;16424
> 2;0.00034233;16428
> 2;0.00034231;16432
> 2;0.000342906;16436
> 2;0.000342528;16440
> 2;0.000360112;16444<--
> 2;0.000343095;16448
> 2;0.000342613;16452
> 2;0.000342823;16456
> 2;0.000342695;16460
> 2;0.000342781;16464
> 2;0.000342082;16468
> 2;0.000342706;16472
> 2;0.000342361;16476
> 2;0.000342908;16480
> 2;0.000358126;16484<--
> 2;0.000342697;16488
> 2;0.000342495;16492
> 2;0.000342806;16496
> 2;0.000342006;16500
> 2;0.000342573;16504
> 2;0.000343114;16508
> 2;0.000343439;16512
> 2;0.000343044;16516
> 2;0.000343867;16520
> 2;0.000357508;16524<--
>
> If somebody could explain me?
>
> Regards
>
> Ulrich David
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
>
--
{+} Jeff Squyres
{+} The Open MPI Project
{+} http://www.open-mpi.org/
|
