On 11/01/2018 08:03 PM, Andres Freund wrote:
> On 2018-11-01 19:57:17 +0100, Tomas Vondra wrote:
>>>> I think that very much depends on how expensive the tasks handled by the
>>>> threads are. It may still be cheaper than a reasonable IPC, and if you
>>>> don't create/destroy threads, that also saves quite a bit of time.
>>>
>>> I'm not following. How can you have a pool *and* threads? Those seem to
>>> be contradictory in PG's architecture? You need full blown IPC with your
>>> proposal afaict?
>>>
>>
>> My suggestion was to create a bgworker, which would then internally
>> allocate and manage a pool of threads. It could then open some sort of
>> IPC (say, as dumb as unix socket). The backends could could then send
>> requests to it, and it would respond to them. Not sure why/how would
>> this contradict PG's architecture?
>
> Because you said "faster than reasonable IPC" - which to me implies that
> you don't do full blown IPC. Which using threads in a bgworker is very
> strongly implying. What you're proposing strongly implies multiple
> context switches just to process a few results. Even before, but
> especially after, spectre that's an expensive proposition.
>
Gah! I meant to wrote "faster with reasonable IPC" - i.e. faster/cheaper
than a solution that would create threads ad-hoc.
My assumption is that the tasks are fairly large, and may take quite a
bit of time to process (say, a couple of seconds?). In which cese the
the extra context switches are not a major issue. But maybe I'm wrong.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services