Josh Berkus <josh@agliodbs.com> writes:
>> Most likely that's the libc implementation of the select()-based sleeps
>> for vacuum_cost_delay. I'm still suspicious that the writes are eating
>> more cost_delay points than you think.
> Tested that. It does look like if I increase vacuum_cost_limit to 10000
> and lower vacuum_cost_page_dirty to 10, it reads 5-7 pages and writes
> 2-3 before each pollsys. The math seems completely wrong on that,
> though -- it should be 50 and 30 pages, or similar.
I think there could be a lot of cost_delay points getting expended
without any effects visible at the level of strace. Maybe try fooling
with vacuum_cost_page_hit and vacuum_cost_page_miss, too.
regards, tom lane