Wes <wespvp@syntegra.com> writes:
> Watching the system as vacuum is running, I can see that we are encountering
> the kswapd/kscand problem in the 2.4.20 kernel. This could very well
> account for the non-linear increase in vacuum time.
Hmm. Looking at the vacuum verbose output you sent me, it's clear that
the bulk of the time is going into scanning a couple of the larger
indexes. On an index that's been growing for awhile, this involves a
highly nonsequential access pattern (it wants to visit the index leaf
pages in sort order, which will not look much like physical order after
a lot of page splits have occurred). I don't know whether that would
tend to set off the kswapd/kscand problems, but just in terms of
physical I/O time it might be annoying. I was going to suggest
REINDEXing those indexes to see if that cuts the vacuum time at all.
regards, tom lane