On Wed, Jan 12, 2005 at 06:37:51PM -0800, Roland Dreier wrote: > This may or may not help on your ia64 systems, but one thing I've > found helpful for getting the highest IPoIB performance is to bind the > HCA's interrupt handlers to one CPU and use taskset to bind the > netperf/netserver processes to a different CPU.
Ok - thanks for the suggestion but it doesn't help. taskset 1 /opt/netperf/netperf -l 60 -H 10.0.0.81 -t TCP_STREAM -- -m 8192 -s 262144 -S 262144 and then manually determining the netserver PID and setting it to various masks. I got: netperf/irq same, netserver/irq same: 1591 Mbps netperf/irq diff, netserver/irq same: 1565 Mbps netperf/irq same, netserver/irq diff: 1507 Mbps netperf/irq diff, netserver/irq diff: 1510 Mbps It's take ~5-10 seconds to locate the PID and set the mask. The last two numbers are too close to draw any conclusions. I didn't expect it to improve since the CPU isn't fully utilized. And it looks like caching is playing a measurable role on the receive side of things. Some prefetching might help...I'll play with that later. thanks, grant _______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
