I've just put enough instrumentation in to get some numbers. On Sep 7, 2010, at 4:34 AM, Martin Lucina wrote:
> Matt, > > [email protected] said: >> I've got about 2,000 threads talking over REQ ports inproc to a >> single XREP port. >> >> Any thoughts as to whether this might be the cause of performance >> hiccups under 2.0.7? I'm polling the XREP and a few other ports in >> my main loop, and have already raised the device to SCHED_RR ... I >> believe the kernel is clocked at 1kHz and running CFS... > > What kind of performance hiccups? Can you be more specific? > My device thread is running at SCHED_RR pri=66, I have approximately 256 client threads, the time from last packet send() on REQ to last device recv() on XREP is averaging around 40ms with a stddev of about 36-40 (10,000 samples) There's limited or no variation reordering poll upcalls (reactors) The other direction (device XREP >> user REQ) is almost always nil. I haven't declared any HWMs so presumably they are infinite. > What platform are you running on? > Linux dev2201.corp.xplusone.com 2.6.18-194.8.1.el5xen #1 SMP Thu Jul 1 19:41:05 EDT 2010 x86_64 x86_64 x86_64 GNU/Linux > -mato > _______________________________________________ > zeromq-dev mailing list > [email protected] > http://lists.zeromq.org/mailman/listinfo/zeromq-dev _______________________________________________ zeromq-dev mailing list [email protected] http://lists.zeromq.org/mailman/listinfo/zeromq-dev
