Re: Impact of checkpoint_segments under continual load conditions

On 7/18/05, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> > The table has 15 columns, 5 indexes (character, inet and timestamp).
> > No foreign keys. The only other thing running on the machine was the
> > application actually DOING the benchmarking, written in Python
> > (psycopg), but it was, according to top, using less than 1% of the
> > CPU. It was just talking through a pipe to a psql prompt to do the
> > COPY.
>
> Sounds pretty plain-vanilla all right.
>
> Are you in a position to try the same benchmark against CVS tip?
> (The nightly snapshot tarball would be plenty close enough.) I'm
> just wondering if the old bgwriter behavior of locking down the
> bufmgr while it examined the ARC/2Q data structures is causing this...

Tom,

It looks like the CVS HEAD is definately "better," but not by a huge
amount. The only difference is I wasn't run autovacuum in the
background (default settings), but I don't think this explains it.
Here's a graph of the differences and density of behavior:

http://blog.amber.org/diagrams/pgsql_copy_803_cvs.png

I can provide the raw data. Each COPY was 500 rows. Note that fsync
is turned off here. Maybe it'd be more stable with it turned on?

Chris
--
| Christopher Petrilli
| petrilli(at)gmail(dot)com