[ofa-general] IPoIB-CM/RC - NAPI patch.
Bernard King-Smith
wombat2 at us.ibm.com
Tue Mar 6 06:14:13 PST 2007
"Erez Strauss" <erezs at voltaire.com> wrote on 03/06/2007 12:51:52 AM:
> Hi Bernie,
>
> Thank you for your reply.
>
> In this case I?m using an application which is using very small
> messages (300Bytes) and I can not change it.
> The focus of the testing is to reduce the latency to minimum during
> high stress, while keeping the application intact.
O.K. you are looking at small packet performance as opposed to peak
bandwidth. In this case NAPI would lower the interrupt rate if TCP doesn't
already coalesce multiple TCP packets into a larger packet over IB.
However, at a packet size of 300 bytes, I would expect that the traffic
actually comes over the UD QP not the RC one.
With your rate of 40000 interrupts per second, if each interrupt was a
single packet you are only looking at 12 MB/s. If you are not using
TCP_NODELAY then your bandwidth may be higher. Even so, NAPI hopefully
with improve the small packet throughput.
>
> Similar synthetic test would be multiple ?iperf ?l 300 ?.? and sum
> of the total bandwidth (or packets per seconds).
>
> Any suggestions for IPoIB-CM/RC are welcome.
> Do you expect the SDP to perform better under the above constrains?
>
> Thanks,
> Erez
> Voltaire Inc.
>
Regards,
Bernie King-Smith
IBM Corporation
Server Group
Cluster System Performance
wombat2 at us.ibm.com (845)433-8483
Tie. 293-8483 or wombat2 on NOTES
"We are not responsible for the world we are born into, only for the world
we leave when we die.
So we have to accept what has gone before us and work to change the only
thing we can,
-- The Future." William Shatner
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20070306/1b530ed3/attachment.html>
More information about the general
mailing list