Hi, On 2017-10-10 13:40:07 -0700, pinker wrote: > and the total number of connections are increasing very fast (but I suppose > it's the symptom not the root cause of cpu load) and exceed max_connections > (1000). Others mentioned already that that's worth improving. > System: > * CentOS Linux release 7.2.1511 (Core) > * Linux 3.10.0-327.36.3.el7.x86_64 #1 SMP Mon Oct 24 16:09:20 UTC 2016 > x86_64 x86_64 x86_64 GNU/Linux Some versions of this kernel have had serious problems with transparent hugepages. I'd try turning that off. I think it defaults to off even in that version, but also make sure zone_reclaim_mode is disabled. > * postgresql95-9.5.5-1PGDG.rhel7.x86_64 > * postgresql95-contrib-9.5.5-1PGDG.rhel7.x86_64 > * postgresql95-docs-9.5.5-1PGDG.rhel7.x86_64 > * postgresql95-libs-9.5.5-1PGDG.rhel7.x86_64 > * postgresql95-server-9.5.5-1PGDG.rhel7.x86_64 > > * 4 sockets/80 cores 9.6 has quite some scalability improvements over 9.5. I don't know whether it's feasible for you to update, but if so, It's worth trying. How about taking perf profile to investigate? > * vm.dirty_background_bytes = 0 > * vm.dirty_background_ratio = 2 > * vm.dirty_bytes = 0 > * vm.dirty_expire_centisecs = 3000 > * vm.dirty_ratio = 20 > * vm.dirty_writeback_centisecs = 500 I'd suggest monitoring /proc/meminfo for the amount of Dirty and Writeback memory, and see whether rapid changes therein coincide with periodds of slowdown. Greetings, Andres Freund -- Sent via pgsql-general mailing list (pgsql-general@xxxxxxxxxxxxxx) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-general
