>8 x 16GB 1600MHz PC3-12800 DDR3 - 128GB total >>shared_buffers=60GB
I would say 60GB is too high when you have 128GB system memory. Try lowering it to shared_buffers=32GB and let the O/S handle more of the work. On Tue, Aug 18, 2015 at 11:49 AM, Jeff Janes <jeff.ja...@gmail.com> wrote: > On Tue, Aug 18, 2015 at 8:01 AM, Michael H <mich...@wemoto.com> wrote: > >> Hi, >> >> I've been tuning our new database server, here's some info... >> >> CentOS Linux release 7.1.1503 (Core) >> 3.10.0-229.11.1.el7.x86_64 >> >> 8 x 16GB 1600MHz PC3-12800 DDR3 - 128GB total >> 2 x AMD Opteron 6386SE 2.8GHz/16-core/140w - 32 cores total >> 4 x 300GB SAS 10k HDD - raid 1+0 configuration >> 1GB FBWC for P-series smart array - cache enabled >> >> I'm using the CentOS provided packages for PostgreSQL >> Version : 9.2.13 >> Release : 1.el7_1 >> >> I'm getting fairly good statistics from this server but after asking for >> some advice I was pointed towards PostgreSQL 9.3 (posix memory management) >> and PostgreSQL 9.4 (pg_replication_slots). >> >> I dropped my original install of 9.2.13 above and went straight to the >> 9.4 from the PostgreSQL repositories. >> > > > How did you get your data from 9.2 to 9.4? Did you run ANALYZE on it > afterwards? > > > >> Are there any known issues with my kernel and PostgreSQL? I found this >> post - >> >> http://www.databasesoup.com/2014/09/why-you-need-to-avoid-linux-kernel-32.html >> >> which states there are known issues up to kernel 3.10.. the reason I ask, >> no matter how small or big a configuration change I make I can't match my >> 9.2.13 install. I'm seeing huge decreases in TPS on all my benchmarks. >> >> for example, 9.2.13, my own extremely heavy SQL file being used here, >> hence the lower TPS... >> >> 32 37.357197 >> 64 34.145088 >> 128 19.682544 >> 256 9.910772 >> 512 5.803358 >> >> compared to 9.4 - exactly the same tests and parameters configured (I >> also started from defaults and tuned up as best I could). >> >> 32 14.982111 >> 64 14.894859 >> 128 14.277631 >> 256 13.679516 >> 512 13.679516 >> > > Pick the query that dropped in performance the most, then run it with > "explain (analyze, buffers)" and with track_io_timing turned on, and > compare this between the servers. Did the plan change, or just the time? > > Cheers, > > Jeff > -- *Melvin Davidson* I reserve the right to fantasize. Whether or not you wish to share my fantasy is entirely up to you.