On Fri, Nov 15, 2013 at 03:25:00PM -0800, Joe Stringer wrote: > When profiling CPU usage in situations involving high numbers of ports, > coverage_clear() was highlighted as a commonly called function. It > appears that it can be quite expensive to access all of the per-thread > coverage counters when threads are constantly waking up. > > This patch makes each thread only do coverage_clear() logic roughly once > per second by introducing per-thread timers. Upcall handler counters may > become less accurate, as these threads may sleep without synchronising > and not wake up for some time. When the main thread is under load at > ~90% CPU, this drops to ~85%. Upcall handler threads sitting at ~2.5% > drop to ~1.5%. > > Signed-off-by: Joe Stringer <joestrin...@nicira.com>
Applied, thanks. > The existing ARP modification slow-path test fails with this patch, as > the upcall handler goes to sleep before synchronising its per-thread > counters, and warping the time does not affect handler threads. This > patch removes the conflicting part of the test. An alternative is to > introduce a coverage_clear_now() function to be called from the upcall > handler, if we are not concerned about the performance difference for > upcall handlers. For now, I decided to just apply your patch. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
