[Beowulf] SCI Socket latency: 2.27 microseconds

Dan Kidger daniel.kidger at quadrics.com
Mon Mar 1 18:47:33 EST 2004


On Monday 01 March 2004 9:38 pm, Bill Broadley wrote:
> > well, Quadrics now claims 1.8 us MPI latency:
> > http://doc.quadrics.com/quadrics/QuadricsHome.nsf/NewsByDate/5D7D7444A42B
> >398280256E44005A31DD
>
> Note the title says "sub 2us" and the body says "close to" 1.8us.

Just ran this:
[dan at opteron0]$ mpicc mping.c -o mping;  prun -N2 ./mping
  1:        0 bytes      1.80 uSec     0.00 MB/s

This is simple bit of MPI: proc 1 posts an MPI_Recv,  proc0 then does a 
MPI_Send, then proc1 does MPI_Send and proc0 an MPI_Recv.
Latency printed is half the round trip averaged over say 1000 passes

This is for Opteron - it seems to have the best PCI-X implimentation we have 
seen. Latency on IA64 is a little higher - say 2.61uSec on one platform I 
have just tried.
   MPI performance has also improved over time as we have tuned the DMA/PIO 
writes,etc. in the device drivers.


> Of more interest (to me) is that further down they say:
>  In the next quarter, Quadrics will announce a series of highly competitive
>  switch configurations making QsNetII more cost-effective for medium
>  sized cluster configuration deployment.

yep - yet to be announced offically - but as you might expect this revolves 
around introducing a wider range of smaller switch chasses and 
configuratiions.   


-- 
Yours,
Daniel.

--------------------------------------------------------------
Dr. Dan Kidger, Quadrics Ltd.      daniel.kidger at quadrics.com
One Bridewell St., Bristol, BS1 2AA, UK         0117 915 5505
----------------------- www.quadrics.com --------------------



_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list