> > > > However, I agree that 100 is less than 200 and queue will be overflown if > > ab > > generates requests faster than the app processes them. Sure ab fires > > things > > quite fast. > > > > I oversimplified my production case. In the presented case it's 1.5ms > > between accept() > > calls. In production requests I see 600ms gaps between accept() calls. > > This > > is what > > brought me here to the list. > > > > Probably I have to dig more into production straces... > > > > What do you think? Ideas? > >
ab and wrk are tool for measuring throughput, not emulating realistic workload. They make 200 connection at once (within very short time rage) if you specified 200 concurrency. I think constant rate load testing tool (e.g. vegeta, wrk2) may help you. And iff you want to emulate production requests, you must use other load simulator tools (e.g. locust). _______________________________________________ uWSGI mailing list uWSGI@lists.unbit.it http://lists.unbit.it/cgi-bin/mailman/listinfo/uwsgi