On Fri, Dec 10, 2004 at 04:42:09PM -0800, Grant Grundler wrote: > I'll dig up the other trivial things with pfmon.
pfmon 3.1 hasn't been released like I'd hoped and thus I may not be able to collect Data EAR like I hoped. I did a q-syscollect run (full output): http://gsyprf3.external.hp.com/apache2-default/openib/q-1321-tcp_stream-0.txt Here are the top offenders: Flat profile of CPU_CYCLES in kernel-cpu0.hist#0: Each histogram sample counts as 1.00034m seconds % time self cumul calls self/call tot/call name 20.48 8.10 8.10 41.0k 198u 198u default_idle 15.02 5.94 14.05 1.83M 3.25u 5.06u mthca_interrupt 14.91 5.90 19.94 17.2M 342n 342n _spin_unlock_irqrestore 4.05 1.60 21.55 12.1M 132n 149n ipt_do_table 3.20 1.27 22.81 7.17M 177n 177n do_csum 2.67 1.05 23.87 7.83M 135n 135n __copy_user 2.26 0.89 24.76 33.9M 26.4n 36.6n local_bh_enable 2.15 0.85 25.61 2.81M 303n 778n tcp_transmit_skb 2.01 0.79 26.41 1.45M 545n 3.98u tcp_sendmsg ... hrm...don't understand the 20% idle. This is a dual CPU system and (this version of) netperf is not multi-threaded. The top 3 only add up to about 50%. I guess I need to see what's being inlined into mthca_interrupt and try to break that down into smaller bits. 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
