Re:Pgagroal: High-performance connection pool for PostgreSQL
by anarazel on Monday November 22, 2021 @02:41PM
(#29310751)
(Although the sibling post pointing to my postgres connection scalability post is right - processes are not the major bottleneck at the moment, and certainly not with < PG 14)
Re:Pgagroal: High-performance connection pool for PostgreSQL
by jpgvm on Monday November 22, 2021 @02:53PM
(#29310889)
Yeah I agree wholeheartedly. What I was trying to get across was that simply s/processes/threads isn't going to make a massive difference to postgres if that was the only change you were to make.
Your analysis seems pretty on-point with what I expected, i.e snapshot scalability and other postgres specific arch being the major bottlenecks for higher connection counts.
Re:Pgagroal: High-performance connection pool for PostgreSQL
by anarazel on Monday November 22, 2021 @03:07PM
(#29311024)
FWIW, a good chunk of the snapshot scalability changes are in PG 14, so the situation has gone a bit further towards s/processes/threads/ being the primary bottleneck. It depends quite a bit on whether you're write-mostly (bottlenecks almost completely unrelated to processes vs threads) or read-mostly. And in the latter case whether it's a context switch heavy workload (processes vs threads may be a significant factor), or not (processes vs threads can be a significant factor, but less likely).
Not really. The page-table overhead of a 100k processes alone is going to a eat up a lot of memory - and you don't have that with threads. And with something like postgres the page tables aren't going to be tiny, even with explicit use of huge pages for applicable allocations. The decrease in TLB hit ratio also is very painful. In a lot of workloads postgres spends way too much time in TLB faults, and that's to a good degree caused by the process model (also caused by linux making it way too hard to map executables with huge pages, itlb faults are a significant issue).
That's not to say that a 1 connection: 1 thread approach is the right thing for a pooler - threads area also don't scale great. I think there are good simplicity arguments to be made for a 1 connection: 1 thread model for a database server itself, but for poolers the benefits seem much smaller.