Re: Re: [COMMITTERS] pgsql: Revert "commit_delay" change; just add comment that we don't hav

On 08/15/2012 11:41 AM, Peter Geoghegan wrote:
> I know that someone is going to point out that in some particularly benchmark,
> they can get another relatively modest increase in throughput (perhaps
> 2%-3%) by splitting the difference between two adjoining millisecond
> integer values. In that scenario, I'd be tempted to point out that
> that increase is quite unlikely to carry over to real-world benefits,
> because the setting is then right on the cusp of where increasing
> commit_delay stops helping throughput and starts hurting it.

You guessed right on that. I just responded to your survey over on
pgsql-performance with two cases where older versions found optimal
performance with commit_delay in the <=10 usec range. Those are all in
the BBWC case that I don't think you've been testing much of yet.

I recall Jignesh Shah reported his seeing that was from slightly better
chunking of writes to disk, with a small but measurable drop in disk I/O
operations (such as IOPS) relative to TPS. The average throughput was
no different; the number of *operations* was smaller though. Less 8K
I/O requests, more 16K+ ones. Like a lot of these situations, adding
some latency to every transactions can make them batch better. And that
can unexpectedly boost throughput enough that net latency is actually
faster. It's similar to how adding input queue latency with a pooler,
limiting active connections, can actually make latency better by
increasing efficiency.

On higher-end storage you can reach a point where IOPS gets high enough
that the per-operation overhead becomes a problem, on top of the usual
"is there enough write throughput?" question. I suspect this situation
might even be more common now, given IOPS issues like this are commonly
highlighted when people do SSD reviews.

I still don't know that it's a widely popular situation. But this
particular use case has been one of the more persistent ones arguing to
keep the parameter around until now. Making sub-microsecond resolution
on the parameter go away would effectively trash it just when it might
get even more useful than before.

--
Greg Smith 2ndQuadrant US greg(at)2ndQuadrant(dot)com Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support www.2ndQuadrant.com