Horton

Re: SLRU optimization - configurable buffer pool and partitioning the SLRU lock

x4mmm@yandex-team.ru

View thread View on postgr.es
From: "Andrey M. Borodin" <x4mmm@yandex-team.ru>
To: Robert Haas <robertmhaas@gmail.com>
Cc: Alvaro Herrera <alvherre@alvh.no-ip.org>, Dilip Kumar <dilipbalaut@gmail.com>, tender wang <tndrwang@gmail.com>, PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Date: 2023-12-18T17:53:43Z
Lists: pgsql-hackers 

> On 18 Dec 2023, at 22:30, Robert Haas <robertmhaas@gmail.com> wrote:
> 
> On Mon, Dec 18, 2023 at 12:04 PM Robert Haas <robertmhaas@gmail.com> wrote:
>> certain sense they are competing for the same job. However, they do
>> aim to alleviate different TYPES of contention: the group XID update
>> stuff should be most valuable when lots of processes are trying to
>> update the same page, and the banks should be most valuable when there
>> is simultaneous access to a bunch of different pages. So I'm not
>> convinced that this patch is a reason to remove the group XID update
>> mechanism, but someone might argue otherwise.
> 
> Hmm, but, on the other hand:
> 
> Currently all readers and writers are competing for the same LWLock.
> But with this change, the readers will (mostly) no longer be competing
> with the writers. So, in theory, that might reduce lock contention
> enough to make the group update mechanism pointless.

One page still accommodates 32K transaction statuses under one lock. It feels like a lot. About 1 second of transactions on a typical installation.

When the group commit was committed did we have a benchmark to estimate efficiency of this technology? Can we repeat that test again?


Best regards, Andrey Borodin.

Commits

  1. Fix zeroing of pg_serial page without SLRU bank lock

  2. Fix misspelled assertions

  3. GUC table: Add description to computed variables

  4. Improve performance of subsystems on top of SLRU

  5. Rename SLRU elements in view pg_stat_slru

  6. Use atomic access for SlruShared->latest_page_number

  7. Split use of SerialSLRULock, creating SerialControlLock

  8. Index SLRUs by 64-bit integers rather than by 32-bit integers

  9. Add a macro templatized hashtable.