On Tue, Nov 22, 2011 at 11:41 PM, Bruce Momjian
<bruce@momjian.us> wrote:
Amitabh Kant wrote:> >
Hard to say. We don't know why this is happening but we are guessing it
is the overhead of managing over one million shared buffers. Please
test and let us know.
I do have a FreeBSD 8.0 server running Postgresql 8.4.9 with 32 GB of RAM (Dual Processor, SAS HD/RAID 10 with BBU for data and SAS HD/RAID 1 for pg_xlog). The current settings set effective cache size to 22 GB (pgtune generated values). The server sees around between 500 to 1200 TPS and chugs along pretty nicely. Sadly it's in production so I am not in a position to run any tests on it. Changed values for postgresql.conf are:
==========================================================
maintenance_work_mem = 1GB # pg_generate_conf wizard 2010-05-09
checkpoint_completion_target = 0.9 # pg_generate_conf wizard 2010-05-09
effective_cache_size = 22GB # pg_generate_conf wizard 2010-05-09
work_mem = 320MB # pg_generate_conf wizard 2010-05-09
wal_buffers = 8MB # pg_generate_conf wizard 2010-05-09
checkpoint_segments = 16 # pg_generate_conf wizard 2010-05-09
shared_buffers = 3840MB # pg_generate_conf wizard 2010-05-09
==========================================================
I do have another server that is to go into production(48 GB RAM, dual processor, Intel 710 SSD for data in RAID 1, SAS HD/RAID 1 for pg_xlog). Apart from running pgbench and bonnie++, is there something else that I should be testing on the new server? Greg's stream scaling seems to be for linux, so not sure if it will work in FreeBSD.