Re: gin index creation performance problems

On Mon, 03 Nov 2008 11:04:45 -0500
Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Ivan Sergio Borgonovo <mail(at)webthatworks(dot)it> writes:
> > maintenance_work_mem is still untouched. What would be a good
> > value to start from?

> GIN index build time is *very* sensitive to maintenance_work_mem.
> Try cranking it up to a couple hundred megabytes and see if that
> helps.

200MB did improve the situation. I was waiting a clue about a
reasonable value from where to start.

But the performance is far from being stable.
It can take few minutes to more than 20min even with this setting.
I can't understand if it is CPU bound or RAM bound.
CPU load is always near 100% while postgresql is using 15% of RAM.
Still I've 52K of swap used...
But anyway the performance is very erratic.

BTW Is maintenance_work_mem set per connection?

While gin index looks appreciably faster (actually it is
lightening fast) for searches I'm considering to revert to gist since
even with 200MB maintenance_work_mem it still look a pain to build
up the index, especially considering the random time required to
build it up.

What puzzle me is that while before increasing maintenance_work
mem it was always terribly slow now there is a huge variation in
rebuilt time with 200MB.
Even
vacuum full;
can be pretty slow (>5min) and still 100% CPU use.

The index is dropped at the beginning and rebuilt at the end inside a
quite busy transaction.
Could it be related?

> Also, if you're on 8.2, I think 8.3 might be faster.

8.3 etch backport.

thanks

--
Ivan Sergio Borgonovo
http://www.webthatworks.it