[knem-devel] (no subject)

Brice Goglin Brice.Goglin at inria.fr
Mon Feb 7 10:54:11 CET 2011


Le 05/02/2011 15:40, neuerrechner at gmx.de a écrit :
> I guess that indeed I/OAT is not enabled as there are no such lines in
> lspci. But I could not find a BIOS switch to enable this although I
> guess the Chipset and CPU should feature it.

I have seen some machines (from HP for instance) where there is no such
switch.

> Okay, I see, so maybe it is not that important to further investigate why this I/OAT thing is not working.
> But the point is, that with large MPI messages, there is still some speedup when using knem? I will make some test with and w/o it...
>   

KNEM uses a single copy (while the usual user-space solution uses 2
copies), so KNEM is reducting the CPU utilization, cache pollution, and
memory usage.

If you run some benchmarks with communication intensive patterns
(ping-ping between multiple pairs of cores, alltoall, ...), you should
see some speedups.

If you run some basic benchmarks with very few messages (ping-pong with
2 processes), you might not see a big difference because your Westmere
processors are very fast at copying and the memory interconnect doesn't
get overloaded when not sending many messages at the same time.

Also note that intersocket transfers are also more improved than
intrasocket because there are no shared caches between cores of
different sockets.

Brice




More information about the knem-devel mailing list