Horton

Re: Improve WALRead() to suck data directly from WAL buffers when possible

Masahiko Sawada <sawada.mshk@gmail.com>

View thread
From: Masahiko Sawada <sawada.mshk@gmail.com>
To: Andres Freund <andres@anarazel.de>
Cc: Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>, Jeff Davis <pgsql@j-davis.com>, Dilip Kumar <dilipbalaut@gmail.com>, Kyotaro Horiguchi <horikyota.ntt@gmail.com>, pgsql-hackers@lists.postgresql.org, SATYANARAYANA NARLAPURAM <satyanarlapuram@gmail.com>
Date: 2023-01-27T06:46:04Z
Lists: pgsql-hackers

Commits

Same data as JSON: GET /api/v1/messages/:b64id/commits the thread's linked commits as JSON, with link sources. API reference →

  1. Add XLogCtl->logInsertResult

  2. Add assert to WALReadFromBuffers().

  3. Read WAL directly from WAL buffers.

  4. Additional write barrier in AdvanceXLInsertBuffer().

  5. Use 64-bit atomics for xlblocks array elements.

  6. Don't trust unvalidated xl_tot_len.

On Fri, Jan 27, 2023 at 3:17 PM Andres Freund <andres@anarazel.de> wrote:
>
> Hi,
>
> On 2023-01-27 14:24:51 +0900, Masahiko Sawada wrote:
> > If I'm understanding this result correctly, it seems to me that your
> > patch works well with the WAL DIO patch (WALDIO vs. WAL DIO & WAL
> > BUFFERS READ), but there seems no visible performance gain with only
> > your patch (HEAD vs. WAL BUFFERS READ). So it seems to me that your
> > patch should be included in the WAL DIO patch rather than applying it
> > alone. Am I missing something?
>
> We already support using DIO for WAL - it's just restricted in a way that
> makes it practically not usable. And the reason for that is precisely that
> walsenders need to read the WAL. See get_sync_bit():
>
>         /*
>          * Optimize writes by bypassing kernel cache with O_DIRECT when using
>          * O_SYNC and O_DSYNC.  But only if archiving and streaming are disabled,
>          * otherwise the archive command or walsender process will read the WAL
>          * soon after writing it, which is guaranteed to cause a physical read if
>          * we bypassed the kernel cache. We also skip the
>          * posix_fadvise(POSIX_FADV_DONTNEED) call in XLogFileClose() for the same
>          * reason.
>          *
>          * Never use O_DIRECT in walreceiver process for similar reasons; the WAL
>          * written by walreceiver is normally read by the startup process soon
>          * after it's written. Also, walreceiver performs unaligned writes, which
>          * don't work with O_DIRECT, so it is required for correctness too.
>          */
>         if (!XLogIsNeeded() && !AmWalReceiverProcess())
>                 o_direct_flag = PG_O_DIRECT;
>
>
> Even if that weren't the case, splitting up bigger commits in incrementally
> committable chunks is a good idea.

Agreed. I was wondering about the fact that the test result doesn't
show things to satisfy the first motivation of this patch, which is to
improve performance by reducing disk I/O and system calls regardless
of the DIO patch. But it makes sense to me that this patch is a part
of the DIO patch series.

I'd like to confirm whether there is any performance regression caused
by this patch in some cases, especially when not using DIO.

Regards,

-- 
Masahiko Sawada
Amazon Web Services: https://aws.amazon.com