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

Peter Geoghegan <peter@2ndquadrant.com> writes:
> On 14 August 2012 21:26, Bruce Momjian <bruce@momjian.us> wrote:
>> Revert "commit_delay" change; just add comment that we don't have
>> a microsecond specification.

> I think that if we eventually decide to change the name of
> commit_delay for 9.3 (you previously suggested that that might be
> revisited), it will be reasonable to have the new GUC in units of
> milliseconds.

Well, the reason why it's like that at all is the thought that values
of less than 1 millisecond might be useful.  Are we prepared to suppose
that that is not and never will be true?

            regards, tom lane

On 15 August 2012 05:15, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Peter Geoghegan <peter@2ndquadrant.com> writes:
>> On 14 August 2012 21:26, Bruce Momjian <bruce@momjian.us> wrote:
>>> Revert "commit_delay" change; just add comment that we don't have
>>> a microsecond specification.
>
>> I think that if we eventually decide to change the name of
>> commit_delay for 9.3 (you previously suggested that that might be
>> revisited), it will be reasonable to have the new GUC in units of
>> milliseconds.
>
> Well, the reason why it's like that at all is the thought that values
> of less than 1 millisecond might be useful.  Are we prepared to suppose
> that that is not and never will be true?

I think that the need for sub-millisecond granularity had more to do
with historic quirks of our implementation. Slight tweaks accidentally
greatly improved throughput, if only for the synthetic benchmark in
question. I personally have not seen any need for a sub-millisecond
granularity when experimenting with the setting on 9.3-devel. However,
I am not sure that sub-millisecond granularity could never be of any
use, in squeezing the last small increase in throughput made possible
by commit_delay. Importantly, feedback as the GUC is tuned is far more
predictable than it was with the prior implementation, so this does
seem quite unimportant.

Why does commit_delay have to be an integer? Can't we devise a way of
manipulating it in units of milliseconds, but have the internal
representation be a double, as with pg_stat_statements' total_time
column? That would allow very fine tuning of the delay. As I've
outlined, I'm not sure that it's worth supporting such fine-tuning
with the new implementation.

--
Peter Geoghegan       http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training and Services

Peter Geoghegan <peter@2ndquadrant.com> writes:
> Why does commit_delay have to be an integer? Can't we devise a way of
> manipulating it in units of milliseconds, but have the internal
> representation be a double, as with pg_stat_statements' total_time
> column?

If you wanted to re-implement all the guc.c logic for supporting
unit-ified values such that it would also work with floats, we could
do that.  It seems like way more mechanism than the problem is worth
however.
        regards, tom lane

On 15 August 2012 14:39, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> If you wanted to re-implement all the guc.c logic for supporting
> unit-ified values such that it would also work with floats, we could
> do that.  It seems like way more mechanism than the problem is worth
> however.

Fair enough.

I'm not quite comfortable recommending a switch to milliseconds if
that implies a loss of sub-millisecond granularity. 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. The
improvement is likely to get lost in the noise in the context of a
real-world application, where for example the actually cost of an
fsync is more variable. I'm just not sure that that's the right
attitude.

-- 
Peter Geoghegan       http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training and Services

Peter Geoghegan <peter@2ndquadrant.com> writes:
> I'm not quite comfortable recommending a switch to milliseconds if
> that implies a loss of sub-millisecond granularity. 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. The
> improvement is likely to get lost in the noise in the context of a
> real-world application, where for example the actually cost of an
> fsync is more variable. I'm just not sure that that's the right
> attitude.

To me it's more about future-proofing.  commit_delay is the only
time-interval setting we've got where reasonable values today are in the
single-digit-millisecond range.  So it seems to me not hard to infer
that in a few years sub-millisecond values will be important, whether or
not there's any real argument for them today.
        regards, tom lane

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@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support www.2ndQuadrant.com