On Sat, Dec 09, 2017 at 10:10:25AM +0100, Pavlos Parissis wrote: > On 09/12/2017 05:01 ?u, Christopher Lane wrote: > > It is plausible/expected that my version upgrade performance goes like > > (about 4K connections, > > long lived and short lived mix, TCP only, no HTTP: > > > > 1.5.12 (nproc 1, old connections causing about 100 old -sf processes to > > linger) uses 100% CPU > > almost all the time, frequently with >1 process. (100%, 75%, 48%, ... ). > > Highest CPU user has > > 2-3K connections. > > > > 1.8.1 (nbproc 4, with hard-stop-after 600s) using like 1%, 3%, 3%, 8% CPU > > Also 2-3K > > connections. > > > > You compare multi-process setup with a single process setup, which isn't a > fair comparison. > Do you see the same the performance increase with 1.8.1, when you configure > it with one process?
Good observation! Also I'd add that concurrent connections alone have no impact on CPU usage but on memory. The only exception is if you're using a low performance poller like poll() instead of epoll or kqueue. This is visible in haproxy -vv and might happen if the previous version was built for a generic OS. Cheers, Willy
