Attached is a set of performance measurements attempting to compare 32-bit
executables against 64-bit executables using PostgreSQL 8.0.0beta3 on an
Itanium box running HPUX B.11.23 with 16GB of RAM, dbc_max_pct = 20%
(meaning a 3.2GB kernel buffer cache), and shmmax = 8gb. I haven't
analyzed the data in much depth yet.
The test script (also attached) was run with the postgresql.conf
settings (also attached), except that max_connections and shared_buffers
were varied as noted across the tests. I used pg_bench to measure TPS
(only numbers excluding connection time are included), and varied the
number of concurrent connections, transactions per connection. I have a log
that shows raw output from pgbench if anyone cares for it, ask and I'll
post it (30KB gzipped).
My own early observations are that the optimal setting for shared_buffers
for 32-bit in these tests is around 8192, while for 64-bit its more like
40K. 64-bit 8.0.0beta3 seems roughly equivalent in performance at the
smaller end of the loads I tested, while at the heavier loads, its
performance exceeded the 32-bit performance by a factor of 2. These
observations and all the test data come without any warranty, either
expressed or implied, whatsoever.
I'd be interested in any further analysis or observations.
Description: application/x-gzip (913 bytes)
pgsql-general by date
|Next:||From: Tom Lane||Date: 2004-09-30 21:48:10|
|Subject: Re: about "pg_dump " without pompt password |
|Previous:||From: Ying Lu||Date: 2004-09-30 21:17:31|
|Subject: about "pg_dump " without pompt password|