[openib-general] Re: Re: [PATCH] IPoIB splitting CQ, increase both send/recv poll NUM_WC & interval

[Shirley Ma]

MIchael,

"Michael S. Tsirkin" <mst at mellanox.co.il> wrote on 04/27/2006 01:14:06 PM:
> Quoting r. Shirley Ma <xma at us.ibm.com>:
> > So far work queue gives very consistent 15% througput increase in my
> > local test with one dual core cpu over mthca.
> What happens to the CPU utilization? And latency?

The CPU utilization were doubled under one dual core cpu.
Finally I found that's the problem tx_ring blocked the sender.

After I tune send/recv queues, I got more than double throughput for 
unidirectional netperf with 10-15% more cpu utilization.

I believe after I apply my other removing tx_ring patch, the performance 
would be better.

> > I am planning to add one more cpu to see the difference.
> 
> And what happens on UP?

Don't have a UP.

The patch needs to be verified on large cluster to see how's packets
out of order. I have tried on 8 cpus, it did good.

Thanks
Shirley Ma
IBM Linux Technology Center
15300 SW Koll Parkway
Beaverton, OR 97006-6063
Phone(Fax): (503) 578-7638

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20060427/6c1ff201/attachment.html>