On Wed, Feb 1, 2017 at 1:17 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> Well, the question even for that case is whether it really costs
> anything. My bet is that it is nearly free when it doesn't help, but
> that could be wrong. My experience running pgbench tests is that
> prewarming all of pgbench_accounts on a scale factor that fits in
> shared_buffers using "dd" took just a few seconds, but when accessing
> the blocks in random order the cache took many minutes to heat up.
And benchmarks like dbt-1 have a pre-warming period added in the test
itself where the user can specify in a number of seconds to linearly
increase the load from 0% to 100%, just for filling in the OS and PG's
cache... This feature would be helpful.
> Now, I assume that this patch sorts the I/O (although I haven't
> checked that) and therefore I expect that the prewarm completes really
> fast. If that's not the case, then that's bad. But if it is the
> case, then it's not really hurting you even if the workload changes
> completely.
Having that working fast would be really nice.
--
Michael