Re: Profiling

OK, my turn.

We have two main servers that are identical, one is the online server, the
other is the hot spare. Their specs:

Dual PIII-750 1.5 Gig ram and dual 18 Gig 10krpm UW SCSI drives.
OS on one drive, postgresql on the other.

Interesting postgresql.conf entries:

max_connections = 128
shared_buffers = 32768
max_fsm_relations = 10000
sort_mem = 2048
vacuum_mem = 8192
cpu_tuple_cost = 0.01
cpu_index_tuple_cost = 0.0001
cpu_operator_cost = 0.05

pgbench -c 4 -t 200 delivers about 240 tps.

Performance is outstanding. This machine runs apache, OpenLDAP, Real
Server 1, as well as Postgresql. All non-database data is stored on a
NAS, so the local drives are only used for swap and postgresql. Average
load is about 40 to 200 reads per minute, with only a handful of writes
per minute (1 to 5 max). Most data is loaded from nightly runs out of
the mainframe and a few other systems for things like company phonebook
and ldap.

My test servers:

Server A: Dual PPro 200 with 256 Meg RAM and 6x4Gig 10kRPM UW SCSI drives
(3 quantum, 3 seagate) and 2x80Gig 7200 RPM IDE drives.

Data is generally stored on the pair of 80 gig drives, because the 4 gig
scsis just aren't big enough. The 80 gig ides are setup as two 40 gig
mirrors (i.e. they're split in half) with the other half used to store
backups and such.

shared_buffers = 5000

pgbench -c 4 -t 200 yields about 80 tps.

Performance is actually quite good, and this is a box we bought in 1997.

Server B: (My workstation) Celeron 1.1GHz, with 512 MEg RAM and a 40 gig
IDE @7200 RPM, and a 17 Gig IDE @5400 RPM.

shared_buffers = 4096

pgbench -c 4 -t 200 yields about 75 tps. Yes, my dual PPro 200 outruns
this box. But then again, my workstation has KDE up and running with
Mozilla, xmms mp3 player going, and a couple other programs running as
well.

All of these boxes are / were heavily tested before deployment, and we
have never had a problem with postgresql on any of them.