On Fri, Mar 20, 2015 at 02:44:05AM +0530, Saurab t wrote: > Hello Willy , > > @Baptiste: > iptables are nor running , irqbalance is running.
clearly irqbalance is the first thing to stop when you're working in web environments or anything related to low latency. You need to pin your interrupts by hand instead. Assign 1 to 4 cores to IRQs, and 1 to #nbproc for haproxy, and ensure that they don't overlap. Also be careful about hyperthreading, if you have enough cores, take care of not using the second thread of each core as they don't scale as well. If you have too few cores, then using the second thread can increase the overall performance by 15-20%. > Will send the 100% CPU > load haproxy stats snapshot, today when load is about to reach peak. BTW, please click right on the stats page and "save as" in order to save the whole HTML page, it's much more exploitable than a snapshot as we can hover over the various places to see the counters. > > @Willy: > Certainly will go ahead with upgrade. > >http://t55696.web-haproxy.webtalks.info/100-cpu-load-t55696.html > as mentioned on the link : > noslice > noepoll > > Can it be useful for us too? If you look closely, it was a workaround for a bug which was fixed after 1.5-dev19. So please don't randomly apply workarounds for bugs that were fixed before the version you're running, otherwise you'll waste your time not knowing what you're observing. > >You're running in server-close mode. Maybe you're having an improperly > >tuned conntrack on the machine which uses lots of CPU once the table is > >full for example. > We have disabled the conntrack on servers. With iptables Not running. Is > there anywhere it need to be taken care of? No, if you have it completely disabled, that's fine. Running "mpstat -P ALL 1 10" during the load will show how the CPU usage spreads over all CPUs and what uses it. That will definitely be useful as well! Willy
