On Fri, 14 Mar 2008, Justin wrote:
> I played with shared_buffer and never saw much of an improvement from
> 100 all the way up to 800 megs moved the checkpoints from 3 to 30 and
> still never saw no movement in the numbers.
Increasing shared_buffers normally improves performance as the size of the
database goes up, but since the pgbench workload is so simple the
operating system will cache it pretty well even if you don't give the
memory directly to PostgreSQL. Also, on Windows large settings for
shared_buffers don't work very well, you might as well keep it in the
100MB range.
> wal_sync_method=fsync
You might get a decent boost in resuls that write data (not the SELECT
ones) by changing
wal_sync_method = open_datasync
which is the default on Windows. The way you've got your RAID controller
setup, this is no more or less safe than using fsync.
> i agree with you, those numbers are terrible i realized after posting i
> had the option -C turned on if i read the option -C correctly it is
> disconnecting and reconnecting between transactions. The way read -C
> option creates the worst case.
In addition to being an odd testing mode, there's an outstanding bug in
how -C results are computed that someone submitted a fix for, but it
hasn't been applied yet. I would suggest forgetting you ever ran that
test.
> number of clients: 10
> number of transactions per client: 10000
> number of transactions actually processed: 100000/100000
> tps = 1768.940935 (including connections establishing)
> number of clients: 40
> number of transactions per client: 10000
> number of transactions actually processed: 400000/400000
> tps = 567.149831 (including connections establishing)
> tps = 568.648692 (excluding connections establishing)
Note how the total number of transactions goes up here, because it's
actually doing clients x requested transcations in total. The 40 client
case is actually doing 4X as many total operations. That also means you
can expect 4X as many checkpoints during that run. It's a longer run like
this second one that you might see some impact by increasing
checkpoint_segments.
To keep comparisons like this more fair, I like to keep the total
transactions constant and just divide that number by the number of clients
to figure out what to set the -t parameter to. 400000 is a good medium
length test, so for that case you'd get
-c 10 -t 40000
-c 40 -t 10000
as the two to compare.
--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD
