Re: Performance question..

On Wed, 11 Jun 2003, Williams, Travis L, NPONS wrote:

> All,
>     I'm looking for ideas on tweaking pgsql.. here is my machine stats
>
> Processor 0 runs at 550 MHz
> Processor 1 runs at 550 MHz
> Page Size : 4096
> Phys Pages: 131072
> Total Physical memory = 536870912 (512MB)

SNIP

> #    Shared Memory Size
> #
> shared_buffers = 128        # 2*max_connections, min 16

WAYYYY too small.  Try 500 to 2000 for starters. Note that bigger isn't
always better, it's about fitting shared_buffers to your usage.  This is
set in 8k blocks, so 1000 is really only about 8 meg.  Bigger servers have
settings as high as 32768 which is 256 Megs.

SNIP

> #    Non-shared Memory Sizes
> #
> #sort_mem = 512             # min 32
> #vacuum_mem = 8192          # min 1024

Try setting your sort mem a little higher.  It's measured in k, so 8192
would be 8 megs.


> #    Write-ahead log (WAL)
> #
> #wal_files = 0 # range 0-64
> #wal_sync_method = fsync   # the default varies across platforms:
> #               # fsync, fdatasync, open_sync, or open_datasync
> #wal_debug = 0             # range 0-16
> #commit_delay = 0          # range 0-100000
> #commit_siblings = 5       # range 1-1000
> #checkpoint_segments = 3   # in logfile segments (16MB each), min 1
> #checkpoint_timeout = 300  # in seconds, range 30-3600
> #fsync = true

If you're doing a lot of writing, look at using more than one WAL file and
putting the pg_xlog directory on another drive.  You have to shutdown the
postmaster, copy over the pg_xlog dir, move the on in $PGDATA out of the
way, and link to the "new" directory then restart the postmaster.

Also, if you're doing lots of writes, setting a higher commit_delay and
commit_siblings can help.

> #effective_cache_size = 1000  # default in 8k pages

If your machine has 512 Meg of ram, you want to see how much
(approximately) is used by the OS as file cache/buffer.  Divide that by 8k
and put that number into effective_cache_size.

> #random_page_cost = 4

For machines with fast RAID subsystems, random_page_cost may need to be
lowered.  somewhere between 1 and 2.  If all your dataset fits in memory,
set it to 1.  I use 1.4 as a setting on my machine with 1.5 gig.

> #cpu_tuple_cost = 0.01
> #cpu_index_tuple_cost = 0.001
> #cpu_operator_cost = 0.0025

SNIP

That's all I can think of.  If you can afford more memory, that would be
your best upgrade right now.