Horton

Re: LogwrtResult contended spinlock

Andres Freund <andres@anarazel.de>

View thread View on postgr.es
From: Andres Freund <andres@anarazel.de>
To: pgsql-hackers@lists.postgresql.org,Alvaro Herrera <alvherre@2ndquadrant.com>,Pg Hackers <pgsql-hackers@lists.postgresql.org>
Cc: Jaime Casanova <jaime.casanova@2ndQuadrant.com>
Date: 2020-08-31T18:29:38Z
Lists: pgsql-hackers 
Hi, 

On August 31, 2020 11:21:56 AM PDT, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:
>Jaime Casanova recently reported a situation where pglogical
>replicating
>from 64 POS sites to a single central (64-core) node, each with two
>replication sets, causes XLog's info_lck to become highly contended
>because of frequently reading LogwrtResult.  We tested the simple fix
>of
>adding a new LWLock that protects LogwrtResult and LogwrtRqst; that
>seems to solve the problem easily enough.
>
>At first I wanted to make the new LWLock cover only LogwrtResult
>proper,
>and leave LogwrtRqst alone.  However on doing it, it seemed that that
>might change the locking protocol in a nontrivial way.  So I decided to
>make it cover both and call it a day.  We did verify that the patch
>solves the reported problem, at any rate.

Wouldn't the better fix here be to allow reading of individual members without a lock? E.g. by wrapping each in a 64bit atomic.

Andres
-- 
Sent from my Android device with K-9 Mail. Please excuse my brevity.

Commits

  1. Remove bogus assertion in pg_atomic_monotonic_advance_u64

  2. Add XLogCtl->logInsertResult

  3. Operate XLogCtl->log{Write,Flush}Result with atomics

  4. Split XLogCtl->LogwrtResult into separate struct members

  5. Introduce atomic read/write functions with full barrier semantics.

  6. Reduce the number of GetFlushRecPtr() calls done by walsenders.

  7. Remove most volatile qualifiers from xlog.c