Обсуждение: Checkpointer write combining

Hi,

The attached patchset implements checkpointer write combining -- which
makes immediate checkpoints at least 20% faster in my tests.
Checkpointer achieves higher write throughput and higher write IOPs
with the patch.

Besides the immediate performance gain with the patchset, we will
eventually need all writers to do write combining if we want to use
direct IO. Additionally, I think the general shape I refactored
BufferSync() into will be useful for AIO-ifying checkpointer.

The patch set has preliminary patches (0001-0004) that implement eager
flushing and write combining for bulkwrites (like COPY FROM). The
functions used to flush a batch of writes for bulkwrites (see 0004)
are reused for the checkpointer. The eager flushing component of this
patch set has been discussed elsewhere [1].

0005 implements a fix for XLogNeedsFlush() when called by checkpointer
during an end-of-crash-recovery checkpoint. I've already started
another thread about this [2], but the patch is required for the patch
set to pass tests.

One outstanding action item is to test to see if there are any
benefits to spread checkpoints.

More on how I measured the performance benefit to immediate checkpoints:

I tuned checkpoint_completion_target, checkpoint_timeout, and min and
max_wal_size to ensure no other checkpoints were initiated.

With 16 GB shared buffers and io_combine_limit 128, I created a 15 GB
table. To get consistent results, I used pg_prewarm to read the table
into shared buffers, issued a checkpoint, then used Bilal's patch [3]
to mark all the buffers as dirty again and issue another checkpoint.
On a fast local SSD, this proved to be a consistent 20%+ speed up
(~6.5 seconds to ~5 seconds).

- Melanie

[1] https://www.postgresql.org/message-id/CAAKRu_Yjn4mvN9NBxtmsCQSGwup45CoA4e05nhR7ADP-v0WCig@mail.gmail.com
[2] https://www.postgresql.org/message-id/CAAKRu_a1vZRZRWO3_jv_X13RYoqLRVipGO0237g5PKzPa2YX6g%40mail.gmail.com
[3] https://www.postgresql.org/message-id/flat/CAN55FZ0h_YoSqqutxV6DES1RW8ig6wcA8CR9rJk358YRMxZFmw%40mail.gmail.com

On Tue, Sep 2, 2025 at 5:10 PM Melanie Plageman
<melanieplageman@gmail.com> wrote:
>
> The attached patchset implements checkpointer write combining -- which
> makes immediate checkpoints at least 20% faster in my tests.
> Checkpointer achieves higher write throughput and higher write IOPs
> with the patch.

These needed a rebase. Attached v2.

- Melanie

Hi,

Thank you for working on this!

On Tue, 9 Sept 2025 at 02:44, Melanie Plageman
<melanieplageman@gmail.com> wrote:
>
> On Tue, Sep 2, 2025 at 5:10 PM Melanie Plageman
> <melanieplageman@gmail.com> wrote:
> >
> > The attached patchset implements checkpointer write combining -- which
> > makes immediate checkpoints at least 20% faster in my tests.
> > Checkpointer achieves higher write throughput and higher write IOPs
> > with the patch.

I did the same benchmark you did and I found it is %50 faster (16
seconds to 8 seconds).

From 053dd9d15416d76ce4b95044d848f51ba13a2d20 Mon Sep 17 00:00:00 2001
From: Melanie Plageman <melanieplageman@gmail.com>
Date: Tue, 2 Sep 2025 11:00:44 -0400
Subject: [PATCH v2 1/9] Refactor goto into for loop in GetVictimBuffer()

@@ -731,6 +741,13 @@ StrategyRejectBuffer(BufferAccessStrategy
strategy, BufferDesc *buf, bool from_r
         strategy->buffers[strategy->current] != BufferDescriptorGetBuffer(buf))
         return false;

+    buf_state = LockBufHdr(buf);
+    lsn = BufferGetLSN(buf);
+    UnlockBufHdr(buf, buf_state);
+
+    if (!XLogNeedsFlush(lsn))
+        return true;

I think this should return false.

I am planning to review the other patches later and this is for the
first patch only.

--
Regards,
Nazir Bilal Yavuz
Microsoft

On Tue, Sep 9, 2025 at 9:27 AM Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
>

Thanks for the review!

> From 053dd9d15416d76ce4b95044d848f51ba13a2d20 Mon Sep 17 00:00:00 2001
> From: Melanie Plageman <melanieplageman@gmail.com>
> Date: Tue, 2 Sep 2025 11:00:44 -0400
> Subject: [PATCH v2 1/9] Refactor goto into for loop in GetVictimBuffer()
>
> @@ -731,6 +741,13 @@ StrategyRejectBuffer(BufferAccessStrategy
> strategy, BufferDesc *buf, bool from_r
>          strategy->buffers[strategy->current] != BufferDescriptorGetBuffer(buf))
>          return false;
>
> +    buf_state = LockBufHdr(buf);
> +    lsn = BufferGetLSN(buf);
> +    UnlockBufHdr(buf, buf_state);
> +
> +    if (!XLogNeedsFlush(lsn))
> +        return true;
>
> I think this should return false.

Oops, you're right. v3 attached with that mistake fixed.

- Melanie

On Tue, Sep 9, 2025 at 9:39 AM Melanie Plageman
<melanieplageman@gmail.com> wrote:
>
> Oops, you're right. v3 attached with that mistake fixed.

One more fix and a bit more cleanup in attached v4.

- Melanie

On Tue, Sep 9, 2025 at 11:16 AM Melanie Plageman
<melanieplageman@gmail.com> wrote:
>
> One more fix and a bit more cleanup in attached v4.

Okay one more version: I updated the thread on eager flushing the
bulkwrite ring [1], and some updates were needed here.

- Melanie

[1] https://www.postgresql.org/message-id/flat/CAAKRu_Yjn4mvN9NBxtmsCQSGwup45CoA4e05nhR7ADP-v0WCig%40mail.gmail.com

On Tue, 2025-09-09 at 13:55 -0400, Melanie Plageman wrote:
> On Tue, Sep 9, 2025 at 11:16 AM Melanie Plageman
> <melanieplageman@gmail.com> wrote:
> >
> > One more fix and a bit more cleanup in attached v4.
>
> Okay one more version: I updated the thread on eager flushing the
> bulkwrite ring [1], and some updates were needed here.

v5-0005 comments:

  * Please update the comment above the code change.
  * The last paragraph in the commit message has a typo: "potentially
update the local copy of min recovery point, when xlog inserts are
*not* allowed", right?
  * Shouldn't the code be consistent between XLogNeedsFlush() and
XLogFlush()? The latter only checks for !XLogInsertAllowed(), whereas
the former also checks for RecoveryInProgress().

I'm still not sure I understand the problem situation this is fixing,
but that's being discussed in another thread.

Regards,
    Jeff Davis

On Wed, Sep 10, 2025 at 4:24 AM Chao Li <li.evan.chao@gmail.com> wrote:
>

Thanks for the review!

For any of your feedback that I simply implemented, I omitted an
inline comment about it. Those changes are included in the attached
v6. My inline replies below are only for feedback requiring more
discussion.

> On Sep 10, 2025, at 01:55, Melanie Plageman <melanieplageman@gmail.com> wrote:
>
> 2 - 0001
> ```
> --- a/src/backend/storage/buffer/freelist.c
> +++ b/src/backend/storage/buffer/freelist.c
>
> + if (XLogNeedsFlush(lsn))
> + {
> + /*
> + * Remove the dirty buffer from the ring; necessary to prevent an
> + * infinite loop if all ring members are dirty.
> + */
> + strategy->buffers[strategy->current] = InvalidBuffer;
> + return true;
> + }
>
> - return true;
> + return false;
>  }
> ```
>
> We can do:
> ```
>          If (!XLogNeedsFlush(lan))
>             Return false
>
>         /* Remove the dirty buffer ….
>          */
>        Return true;
> }
> ```

This would make the order of evaluation the same as master but I
actually prefer it this way because then we only take the buffer
header spinlock if there is a chance we will reject the buffer (e.g.
we don't need to examine it for strategies except BAS_BULKREAD)

> 4 - 0002
> ```
> - /* OK, do the I/O */
> - FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
> - LWLockRelease(content_lock);
> -
> - ScheduleBufferTagForWriteback(&BackendWritebackContext, io_context,
> -  &buf_hdr->tag);
> + CleanVictimBuffer(buf_hdr, &buf_state, from_ring, io_context);
> ```
> I saw CleanVictimBuffer() will get content_lock from bufdesc and release it, but it makes the code hard to
understand.Readers might be confused that why content_lock is not released after CleanVictimBuffer() without further
readingCleanVictimBuffer(). 
>
> I’d suggest pass content_lock to CleanVictimBuffer() as a parameter, which gives a clear hint that
CleanVictimBuffer()will release the lock. 

I think for this specific patch in the set your idea makes sense.
However, in the later patch to do write combining, I release the
content locks for the batch in CompleteWriteBatchIO() and having the
start buffer's lock separate as a parameter would force me to have a
special case handling this.

I've added a comment to both CleanVictimBuffer() and its caller
specifying that the lock must be held and that it will be released
inside CleanVictimBuffer.

> 5 - 0002
> ```
>   * disastrous system-wide consequences.  To make sure that can't happen,
>   * skip the flush if the buffer isn't permanent.
>   */
> - if (buf_state & BM_PERMANENT)
> - XLogFlush(recptr);
> + if (!XLogRecPtrIsInvalid(buffer_lsn))
> + XLogFlush(buffer_lsn);
> ```
>
> Why this check is changed? Should the comment be updated accordingly as it says “if the buffer isn’t permanent”,
whichreflects to the old code. 

It's changed because I split the logic for flushing to that LSN and
determining the LSN across the Prepare and Do functions. This is
needed because when we do batches, we want to flush to the max LSN
across all buffers in the batch.

I check if the buffer is BM_PERMANENT in PrepareFlushBuffer(). You
make a good point about my comment, though. I've moved it to
PrepareFlushBuffer() and updated it.

> 8 - 0003
> ```
> bool
> +strategy_supports_eager_flush(BufferAccessStrategy strategy)
> ```
>
> This function is only used in bufmgr.c, can we move it there and make it static?

BufferAccessStrategyData is opaque to bufmgr.c. Only freelist.c can
access it. I agree it is gross that I have these helpers and functions
that would otherwise be static, though.

> 10 - 0004
> ```
> --- a/src/backend/storage/buffer/bufmgr.c
> +++ b/src/backend/storage/buffer/bufmgr.c
>
> + limit = Min(max_batch_size, limit);
> ```
>
> Do we need to check max_batch_size should be less than (MAX_IO_COMBINE_LIMIT-1)? Because BufWriteBatch.bufdescs is
definedwith length of MAX_IO_COMBINE_LIMIT, and the first place has been used to store “start”. 

I assert that in StrategyMaxWriteBatchSize(). io_combine_limit is not
allowed to exceed MAX_IO_COMBINE_LIMIT, so it shouldn't happen anyway,
since we are capping ourselves at io_combine_limit. Or is that your
point?

> 11 - 0004
> ```
> --- a/src/backend/storage/buffer/bufmgr.c
> +++ b/src/backend/storage/buffer/bufmgr.c
>
> + for (batch->n = 1; batch->n < limit; batch->n++)
> + {
> + Buffer bufnum;
> +
> + if ((bufnum = StrategySweepNextBuffer(strategy)) == InvalidBuffer)
> + break;
> ```
>
> Is sweep next buffer right next to start? If yes, can we assert that? But my guess is no, if my guess is true, then
isit possible that bufnum meets start? If that’s true, then we should check next buffer doesn’t equal to start. 

Ah, great point. I didn't think about this. Our sweep will always
start right after the start buffer, but then if it goes all the way
around, it will "lap" the start buffer. Because of this and because I
think it is weird to have the sweep variables in the
BufferAccessStrategy object, I've changed my approach in attached v6.
I set sweep_end to be the start block in the batch and then pass
around a sweep cursor variable. Hitting sweep_end is the termination
condition.

> 12 - 0004
> ```
> @@ -4306,19 +4370,22 @@ CleanVictimBuffer(BufferAccessStrategy strategy,
>
>   if (from_ring && strategy_supports_eager_flush(strategy))
>   {
> + uint32 max_batch_size = max_write_batch_size_for_strategy(strategy);
> ```
>
> I think max_batch_size can be attribute of strategy and set it when creating a strategy, so that we don’t need to
calculatein every round of clean. 

Actually, the max pin limit can change quite frequently. See
GetAdditionalPinLimit()'s usage in read stream code. If the query is
pinning other buffers in another part of the query, it can change our
limit.

I'm not sure if I should call GetAdditionalPinLImit() for each batch
or for each run of batches (like in StrategyMaxWriteBatchSize()).
Currently, I call it for each batch (in FindFlushAdjacents()). The
read stream calls it pretty frequently (each
read_stream_start_pending_read()). But, in the batch flush case,
nothing could change between batches in a run of batches. So maybe I
should move it up and out and make it per run of batches...

> 13 - 0004
> ```
> +void
> +CompleteWriteBatchIO(BufWriteBatch *batch, IOContext io_context,
> + WritebackContext *wb_context)
> +{
> + ErrorContextCallback errcallback =
> + {
> + .callback = shared_buffer_write_error_callback,
> + .previous = error_context_stack,
> + };
> +
> + error_context_stack = &errcallback;
> + pgBufferUsage.shared_blks_written += batch->n;
> ```
>
> Should we only increase shared_blks_written only after the loop of write-back is done?

On master, FlushBuffer() does it after smgrwrite() (before writeback).
I think pgBufferUsage is mainly used in EXPLAIN (also
pg_stat_statements) which won't be used until the end of the query and
won't be displayed if we error out.

> 14 - 0004
> ```
> --- a/src/backend/storage/buffer/freelist.c
> +++ b/src/backend/storage/buffer/freelist.c
>
> +uint32
> +max_write_batch_size_for_strategy(BufferAccessStrategy strategy)
> ```
>
> I think this function can be moved to bufmgr.c and make it static.

This technically could be moved, but it is a function giving you
information about a strategy which seemed to fit better in freelist.c.

> 18 - 0007
> ```
> --- a/src/backend/storage/buffer/bufmgr.c
> +++ b/src/backend/storage/buffer/bufmgr.c
>
> + max_batch_size = checkpointer_max_batch_size();
> ```
>
> Look like we don’t need to calculate max_batch_size in the for loop.

I don't think it's in the for loop.

- Melanie