pg.ddx.io pgsql-committers@postgresql.org mailing list archive
help / color / mirror / Atom feedpgsql: Use pre-fetching for ANALYZE
4+ messages / 2 participants
[nested] [flat]
* pgsql: Use pre-fetching for ANALYZE
@ 2021-03-16 18:48 Stephen Frost <sfrost@snowman.net>
0 siblings, 1 reply; 4+ messages in thread
From: Stephen Frost @ 2021-03-16 18:48 UTC (permalink / raw)
To: pgsql-committers@lists.postgresql.org
Use pre-fetching for ANALYZE
When we have posix_fadvise() available, we can improve the performance
of an ANALYZE by quite a bit by using it to inform the kernel of the
blocks that we're going to be asking for. Similar to bitmap index
scans, the number of buffers pre-fetched is based off of the
maintenance_io_concurrency setting (for the particular tablespace or,
if not set, globally, via get_tablespace_maintenance_io_concurrency()).
Reviewed-By: Heikki Linnakangas, Tomas Vondra
Discussion: https://www.postgresql.org/message-id/VI1PR0701MB69603A433348EDCF783C6ECBF6EF0%40VI1PR0701MB6960.eur...
Branch
------
master
Details
-------
https://git.postgresql.org/pg/commitdiff/c6fc50cb40285141fad401321ae21becbaea1c59
Modified Files
--------------
src/backend/commands/analyze.c | 73 ++++++++++++++++++++++++++++++++++++++++--
1 file changed, 71 insertions(+), 2 deletions(-)
^ permalink raw reply [nested|flat] 4+ messages in thread
* Re: pgsql: Use pre-fetching for ANALYZE
@ 2022-06-03 02:30 Andres Freund <andres@anarazel.de>
parent: Stephen Frost <sfrost@snowman.net>
0 siblings, 1 reply; 4+ messages in thread
From: Andres Freund @ 2022-06-03 02:30 UTC (permalink / raw)
To: Stephen Frost <sfrost@snowman.net>; pgsql-hackers@postgresql.org
Hi,
On 2021-03-16 18:48:08 +0000, Stephen Frost wrote:
> Use pre-fetching for ANALYZE
>
> When we have posix_fadvise() available, we can improve the performance
> of an ANALYZE by quite a bit by using it to inform the kernel of the
> blocks that we're going to be asking for. Similar to bitmap index
> scans, the number of buffers pre-fetched is based off of the
> maintenance_io_concurrency setting (for the particular tablespace or,
> if not set, globally, via get_tablespace_maintenance_io_concurrency()).
I just looked at this as part of debugging a crash / hang in the AIO patch.
The code does:
block_accepted = table_scan_analyze_next_block(scan, targblock, vac_strategy);
#ifdef USE_PREFETCH
/*
* When pre-fetching, after we get a block, tell the kernel about the
* next one we will want, if there's any left.
*
* We want to do this even if the table_scan_analyze_next_block() call
* above decides against analyzing the block it picked.
*/
if (prefetch_maximum && prefetch_targblock != InvalidBlockNumber)
PrefetchBuffer(scan->rs_rd, MAIN_FORKNUM, prefetch_targblock);
#endif
I.e. we lock a buffer and *then* we prefetch another buffer. That seems like a
quite bad idea to me. Why are we doing IO while holding a content lock, if we
can avoid it?
Greetings,
Andres Freund
^ permalink raw reply [nested|flat] 4+ messages in thread
* Re: pgsql: Use pre-fetching for ANALYZE
@ 2022-06-03 02:35 Andres Freund <andres@anarazel.de>
parent: Andres Freund <andres@anarazel.de>
0 siblings, 1 reply; 4+ messages in thread
From: Andres Freund @ 2022-06-03 02:35 UTC (permalink / raw)
To: Stephen Frost <sfrost@snowman.net>; pgsql-hackers@postgresql.org
Hi,
On 2022-06-02 19:30:16 -0700, Andres Freund wrote:
> On 2021-03-16 18:48:08 +0000, Stephen Frost wrote:
> > Use pre-fetching for ANALYZE
> >
> > When we have posix_fadvise() available, we can improve the performance
> > of an ANALYZE by quite a bit by using it to inform the kernel of the
> > blocks that we're going to be asking for. Similar to bitmap index
> > scans, the number of buffers pre-fetched is based off of the
> > maintenance_io_concurrency setting (for the particular tablespace or,
> > if not set, globally, via get_tablespace_maintenance_io_concurrency()).
>
> I just looked at this as part of debugging a crash / hang in the AIO patch.
>
> The code does:
>
> block_accepted = table_scan_analyze_next_block(scan, targblock, vac_strategy);
>
> #ifdef USE_PREFETCH
>
> /*
> * When pre-fetching, after we get a block, tell the kernel about the
> * next one we will want, if there's any left.
> *
> * We want to do this even if the table_scan_analyze_next_block() call
> * above decides against analyzing the block it picked.
> */
> if (prefetch_maximum && prefetch_targblock != InvalidBlockNumber)
> PrefetchBuffer(scan->rs_rd, MAIN_FORKNUM, prefetch_targblock);
> #endif
>
> I.e. we lock a buffer and *then* we prefetch another buffer. That seems like a
> quite bad idea to me. Why are we doing IO while holding a content lock, if we
> can avoid it?
It also seems decidedly not great from a layering POV to do the IO in
analyze.c. There's no guarantee that the tableam maps blocks in a way that's
compatible with PrefetchBuffer(). Yes, the bitmap heap scan code does
something similar, but a) that is opt in by the AM, b) there's a comment
saying it's quite crufty and should be fixed.
Greetings,
Andres Freund
^ permalink raw reply [nested|flat] 4+ messages in thread
* Re: pgsql: Use pre-fetching for ANALYZE
@ 2022-06-06 18:52 Stephen Frost <sfrost@snowman.net>
parent: Andres Freund <andres@anarazel.de>
0 siblings, 0 replies; 4+ messages in thread
From: Stephen Frost @ 2022-06-06 18:52 UTC (permalink / raw)
To: Andres Freund <andres@anarazel.de>; +Cc: pgsql-hackers@postgresql.org
Greetings,
* Andres Freund (andres@anarazel.de) wrote:
> On 2022-06-02 19:30:16 -0700, Andres Freund wrote:
> > On 2021-03-16 18:48:08 +0000, Stephen Frost wrote:
> > > Use pre-fetching for ANALYZE
> > >
> > > When we have posix_fadvise() available, we can improve the performance
> > > of an ANALYZE by quite a bit by using it to inform the kernel of the
> > > blocks that we're going to be asking for. Similar to bitmap index
> > > scans, the number of buffers pre-fetched is based off of the
> > > maintenance_io_concurrency setting (for the particular tablespace or,
> > > if not set, globally, via get_tablespace_maintenance_io_concurrency()).
> >
> > I just looked at this as part of debugging a crash / hang in the AIO patch.
> >
> > The code does:
> >
> > block_accepted = table_scan_analyze_next_block(scan, targblock, vac_strategy);
> >
> > #ifdef USE_PREFETCH
> >
> > /*
> > * When pre-fetching, after we get a block, tell the kernel about the
> > * next one we will want, if there's any left.
> > *
> > * We want to do this even if the table_scan_analyze_next_block() call
> > * above decides against analyzing the block it picked.
> > */
> > if (prefetch_maximum && prefetch_targblock != InvalidBlockNumber)
> > PrefetchBuffer(scan->rs_rd, MAIN_FORKNUM, prefetch_targblock);
> > #endif
> >
> > I.e. we lock a buffer and *then* we prefetch another buffer. That seems like a
> > quite bad idea to me. Why are we doing IO while holding a content lock, if we
> > can avoid it?
At the end, we're doing a posix_fadvise() which is a kernel call but
hopefully wouldn't do actual IO when we call it. Still, agreed that
it'd be better to do that without holding locks and no objection to
making such a change.
> It also seems decidedly not great from a layering POV to do the IO in
> analyze.c. There's no guarantee that the tableam maps blocks in a way that's
> compatible with PrefetchBuffer(). Yes, the bitmap heap scan code does
> something similar, but a) that is opt in by the AM, b) there's a comment
> saying it's quite crufty and should be fixed.
Certainly open to suggestions. Are you thinking it'd make sense to add
a 'prefetch_block' method to TableAmRoutine? Or did you have another
thought?
Thanks!
Stephen
Attachments:
signature.asc (application/pgp-signature, 818B) download
^ permalink raw reply [nested|flat] 4+ messages in thread
end of thread, other threads:[~2022-06-06 18:52 UTC | newest]
Thread overview: 4+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2021-03-16 18:48 pgsql: Use pre-fetching for ANALYZE Stephen Frost <sfrost@snowman.net>
2022-06-03 02:30 ` Andres Freund <andres@anarazel.de>
2022-06-03 02:35 ` Andres Freund <andres@anarazel.de>
2022-06-06 18:52 ` Stephen Frost <sfrost@snowman.net>
This inbox is served by DDX for PostgreSQL; see mirroring instructions
for how to clone and mirror all data and code used for this inbox