[ofa-general] Re: [openib-general] IPOIB NAPI

[Shirley Ma]

Hello Roland,

      I would like to continue this discussion. We have a severe customer
issue on our HS21. A simple netperf TCP_STREAM test shows that 40-50% UDP
receiving errors for 5300MB/s throughput no matter how big the UDP buffer
size is. We have seen heavy interrupt rates on two CPUs. After we bind
netserver to less busy CPU, the receiving errors are totally gone, we are
able to gain around 680MB/s UDP throughput. But that doesn't work for
customer's UDP application, now their UDP applications are not performing
well in this cluster.

      I thought Mellanox does IRQ affinity, why it used 2 CPUs for one link
here?
      I do think we need the missed event patch in upper stream and modify
the UPLs to use event mode to reduce interrupt rates independent of NAPI.

Thanks
Shirley Ma
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20070313/1c380c1b/attachment.html>