walwriter interacts quite badly with synchronous_commit=off

On 25/10/2023 02:09, Andres Freund wrote:

Because of the inherent delay between the checks of XLogCtl->WalWriterSleeping
and Latch->is_set, we also sometimes end up with multiple processes signalling
walwriter, which can be bad, because it increases the likelihood that some of
the signals may be received when we are already holding WALWriteLock, delaying
its release...

That can only happen when walwriter has just come out of "hibernation",
ie. when the system has been idle for a while. So probably not a big
deal in practice.

I think the most important optimization we need is to have
XLogSetAsyncXactLSN() only wake up if there is a certain amount of unflushed
WAL. Unless walsender is hibernating, walsender will wake up on its own after
wal_writer_delay. I think we can just reuse WalWriterFlushAfter for this.

E.g. a condition like
if (WriteRqstPtr <= LogwrtResult.Write + WalWriterFlushAfter * XLOG_BLCKSZ)
return;
drastically cuts down on the amount of wakeups, without - I think - loosing
guarantees around synchronous_commit=off.

In the patch, you actually did:

+		if (WriteRqstPtr <= LogwrtResult.Flush + WalWriterFlushAfter * XLOG_BLCKSZ)
+			return;

It means that you never wake up the walwriter to merely *write* the WAL.
You only wake it up if it's going to also fsync() it. I think that's
correct and appropriate, but it took me a while to reach that conclusion:

It might be beneficial to wake up the walwriter just to perform a
write(), to offload that work from the backend. And walwriter will
actually also perform an fsync() after finishing the current segment, so
it would make sense to also wake it up when 'asyncXactLSN' crosses a
segment boundary. However, if those extra wakeups make sense, they would
also make sense when there are no asynchronous commits involved.
Therefore those extra wakeups should be done elsewhere, perhaps
somewhere around AdvanceXLInsertBuffer(). The condition you have in the
patch is appropriate for XLogSetAsyncXactLSN().

Another reason to write the WAL aggressively, even if you don't flush
it, would be to reduce the number of lost transactions on a segfault.
But we don't give any guarantees on that, and even with the current
aggressive logic, we only write when a page is full so you're anyway
going to lose the last partial page.

It also took me a while to convince myself that this calculation matches
the calculation that XLogBackgroundFlush() uses to determine whether it
needs to flush or not. XLogBackgroundFlush() first divides the request
and result with XLOG_BLCKSZ and then compares the number of blocks,
whereas here you perform the calculation in bytes. I think the result is
the same, but to make it more clear, let's do it the same way in both
places.

See attached. It's the same logic as in your patch, just formulatd more
clearly IMHO.

Because of the frequent wakeups, we do something else that's not smart: We
write out 8k blocks individually, many times a second. Often thousands of
8k pwrites a second.

Even with this patch, when I bumped up wal_writer_delay to 2 so that the
wal writer gets woken up by the async commits rather than the timeout,
the write pattern is a bit silly:

$ strace -p 1099926 # walwriter
strace: Process 1099926 attached
epoll_wait(10, [{events=EPOLLIN, data={u32=3704011232,
u64=94261056289248}}], 1, 1991) = 1
read(3,
"\27\0\0\0\0\0\0\0\0\0\0\0<\312\20\0\350\3\0\0\0\0\0\0\0\0\0\0\0\0\0\0"...,
1024) = 128
pwrite64(5,
"\24\321\5\0\1\0\0\0\0\300\0\373\5\0\0\0+\0\0\0\0\0\0\0\0\n\0\0n\276\242\305"...,
1007616, 49152) = 1007616
fdatasync(5) = 0
pwrite64(5, "\24\321\5\0\1\0\0\0\0
\20\373\5\0\0\0003\0\0\0\0\0\0\0\320\37\20\373\5\0\0\0"..., 16384,
1056768) = 16384
epoll_wait(10, [{events=EPOLLIN, data={u32=3704011232,
u64=94261056289248}}], 1, 2000) = 1
read(3,
"\27\0\0\0\0\0\0\0\0\0\0\0<\312\20\0\350\3\0\0\0\0\0\0\0\0\0\0\0\0\0\0"...,
1024) = 128
pwrite64(5,
"\24\321\5\0\1\0\0\0\0`\20\373\5\0\0\0+\0\0\0\0\0\0\0\0\n\0\0\5~\23\261"...,
1040384, 1073152) = 1040384
fdatasync(5) = 0
pwrite64(5, "\24\321\4\0\1\0\0\0\0@
\373\5\0\0\0\0\0\0\0\0\0\0\0;\0\0\0\264'\246\3"..., 16384, 2113536) = 16384
epoll_wait(10, [{events=EPOLLIN, data={u32=3704011232,
u64=94261056289248}}], 1, 2000) = 1
read(3,
"\27\0\0\0\0\0\0\0\0\0\0\0<\312\20\0\350\3\0\0\0\0\0\0\0\0\0\0\0\0\0\0"...,
1024) = 128
pwrite64(5, "\24\321\5\0\1\0\0\0\0\200
\373\5\0\0\0003\0\0\0\0\0\0\0\320\177 \373\5\0\0\0"..., 1040384,
2129920) = 1040384
fdatasync(5) = 0

In each cycle, the wal writer writes a full 1 MB chunk
(wal_writer_flush_after = '1MB'), flushes it, and then perform a smaller
write before going to sleep.

Those smaller writes seem a bit silly. But I think it's fine.

More throughput for less CPU, seems neat :)

Indeed, impressive speedup from such a small patch!

I'm not addressing that here, but I think we also have the opposite behaviour
- we're not waking up walwriter often enough. E.g. if you have lots of bulk
dataloads, walwriter will just wake up once per wal_writer_delay, leading to
most of the work being done by backends. We should probably wake walsender at
the end of XLogInsertRecord() if there is sufficient outstanding WAL.

Right, that's basically the same issue that I reasoned through above. I
did some quick testing with a few different settings of wal_buffers,
wal_writer_flush_after and wal_writer_delay, to try to see that effect.
But I was not able to find a case where it makes a difference.

--
Heikki Linnakangas
Neon (https://neon.tech)

Hi,

On 2023-10-25 12:17:03 +0300, Heikki Linnakangas wrote:

On 25/10/2023 02:09, Andres Freund wrote:

Because of the inherent delay between the checks of XLogCtl->WalWriterSleeping
and Latch->is_set, we also sometimes end up with multiple processes signalling
walwriter, which can be bad, because it increases the likelihood that some of
the signals may be received when we are already holding WALWriteLock, delaying
its release...

That can only happen when walwriter has just come out of "hibernation", ie.
when the system has been idle for a while. So probably not a big deal in
practice.

Maybe I am missing something here - why can this only happen when hibernating?
Even outside of that two backends can decide that that they need to wake up
walwriter?

We could prevent that, by updating state when requesting walwriter to be woken
up. But with the changes we're discussing below, that should be rare.

I think the most important optimization we need is to have
XLogSetAsyncXactLSN() only wake up if there is a certain amount of unflushed
WAL. Unless walsender is hibernating, walsender will wake up on its own after
wal_writer_delay. I think we can just reuse WalWriterFlushAfter for this.

E.g. a condition like
if (WriteRqstPtr <= LogwrtResult.Write + WalWriterFlushAfter * XLOG_BLCKSZ)
return;
drastically cuts down on the amount of wakeups, without - I think - loosing
guarantees around synchronous_commit=off.

In the patch, you actually did:

+		if (WriteRqstPtr <= LogwrtResult.Flush + WalWriterFlushAfter * XLOG_BLCKSZ)
+			return;

It means that you never wake up the walwriter to merely *write* the WAL. You
only wake it up if it's going to also fsync() it. I think that's correct and
appropriate, but it took me a while to reach that conclusion:

Yea, after writing the email I got worried that just looking at Write would
perhaps lead to not flushing data soon enough...

It might be beneficial to wake up the walwriter just to perform a write(),
to offload that work from the backend. And walwriter will actually also
perform an fsync() after finishing the current segment, so it would make
sense to also wake it up when 'asyncXactLSN' crosses a segment boundary.
However, if those extra wakeups make sense, they would also make sense when
there are no asynchronous commits involved. Therefore those extra wakeups
should be done elsewhere, perhaps somewhere around AdvanceXLInsertBuffer().
The condition you have in the patch is appropriate for
XLogSetAsyncXactLSN().

Yea. I agree we should wake up walsender in other situations too...

Another reason to write the WAL aggressively, even if you don't flush it,
would be to reduce the number of lost transactions on a segfault. But we
don't give any guarantees on that, and even with the current aggressive
logic, we only write when a page is full so you're anyway going to lose the
last partial page.

Wal writer does end up writing the trailing partially filled page during the
next wal_writer_delay cycle.

It also took me a while to convince myself that this calculation matches the
calculation that XLogBackgroundFlush() uses to determine whether it needs to
flush or not. XLogBackgroundFlush() first divides the request and result
with XLOG_BLCKSZ and then compares the number of blocks, whereas here you
perform the calculation in bytes. I think the result is the same, but to
make it more clear, let's do it the same way in both places.

See attached. It's the same logic as in your patch, just formulatd more
clearly IMHO.

Yep, makes sense!

Because of the frequent wakeups, we do something else that's not smart: We
write out 8k blocks individually, many times a second. Often thousands of
8k pwrites a second.

Even with this patch, when I bumped up wal_writer_delay to 2 so that the wal
writer gets woken up by the async commits rather than the timeout, the write
pattern is a bit silly:

$ strace -p 1099926 # walwriter
strace: Process 1099926 attached
epoll_wait(10, [{events=EPOLLIN, data={u32=3704011232,
u64=94261056289248}}], 1, 1991) = 1
read(3,
"\27\0\0\0\0\0\0\0\0\0\0\0<\312\20\0\350\3\0\0\0\0\0\0\0\0\0\0\0\0\0\0"...,
1024) = 128
pwrite64(5, "\24\321\5\0\1\0\0\0\0\300\0\373\5\0\0\0+\0\0\0\0\0\0\0\0\n\0\0n\276\242\305"...,
1007616, 49152) = 1007616
fdatasync(5) = 0
pwrite64(5, "\24\321\5\0\1\0\0\0\0
\20\373\5\0\0\0003\0\0\0\0\0\0\0\320\37\20\373\5\0\0\0"..., 16384, 1056768)
= 16384
epoll_wait(10, [{events=EPOLLIN, data={u32=3704011232,
u64=94261056289248}}], 1, 2000) = 1
read(3,
"\27\0\0\0\0\0\0\0\0\0\0\0<\312\20\0\350\3\0\0\0\0\0\0\0\0\0\0\0\0\0\0"...,
1024) = 128
pwrite64(5,
"\24\321\5\0\1\0\0\0\0`\20\373\5\0\0\0+\0\0\0\0\0\0\0\0\n\0\0\5~\23\261"...,
1040384, 1073152) = 1040384
fdatasync(5) = 0
pwrite64(5, "\24\321\4\0\1\0\0\0\0@
\373\5\0\0\0\0\0\0\0\0\0\0\0;\0\0\0\264'\246\3"..., 16384, 2113536) = 16384
epoll_wait(10, [{events=EPOLLIN, data={u32=3704011232,
u64=94261056289248}}], 1, 2000) = 1
read(3,
"\27\0\0\0\0\0\0\0\0\0\0\0<\312\20\0\350\3\0\0\0\0\0\0\0\0\0\0\0\0\0\0"...,
1024) = 128
pwrite64(5, "\24\321\5\0\1\0\0\0\0\200 \373\5\0\0\0003\0\0\0\0\0\0\0\320\177
\373\5\0\0\0"..., 1040384, 2129920) = 1040384
fdatasync(5) = 0

In each cycle, the wal writer writes a full 1 MB chunk
(wal_writer_flush_after = '1MB'), flushes it, and then perform a smaller
write before going to sleep.

I think that's actually somewhat sane - we write out the partial page in the
subsequent cycle. That won't happen if the page isn't partially filled or
doesn't have an async commit on it.

I think we end up with somewhat bogus write patterns in other cases still, but
that's really more an issue in XLogBackgroundFlush() and thus deserves a
separate patch/thread.

I'm not addressing that here, but I think we also have the opposite behaviour
- we're not waking up walwriter often enough. E.g. if you have lots of bulk
dataloads, walwriter will just wake up once per wal_writer_delay, leading to
most of the work being done by backends. We should probably wake walsender at
the end of XLogInsertRecord() if there is sufficient outstanding WAL.

Right, that's basically the same issue that I reasoned through above. I did
some quick testing with a few different settings of wal_buffers,
wal_writer_flush_after and wal_writer_delay, to try to see that effect. But
I was not able to find a case where it makes a difference.

I think in the right set of circumstances it can make quite a bit of
difference. E.g. I bulk load 3GB of data in a cluster with s_b 1GB. Then I
checkpoint and VACUUM FREEZE it. With wal_writer_delay=1ms that's
considerably faster (5.4s) than with wal_writer_delay=2s (8.3s) or even the
default 200ms (7.9s), because a fast walwriter makes it much more likely that
vacuum won't need to wait for an xlog flush before replacing a buffer in the
strategy ring.

I think improving this logic would be quite worthwhile!

Another benefit of triggering wakeups based on the amount of outstanding
writes would be that we could increase wal_writer_delay substantially (with
perhaps some adjustment for the partial-trailing-page-with-async-commit case),
reducing power usage. It's imo pretty silly that we have wal writer wake up
regularly, if it just writes once every few seconds.

Greetings,

Andres Freund

On 25/10/2023 21:59, Andres Freund wrote:

On 2023-10-25 12:17:03 +0300, Heikki Linnakangas wrote:

On 25/10/2023 02:09, Andres Freund wrote:

Because of the inherent delay between the checks of XLogCtl->WalWriterSleeping
and Latch->is_set, we also sometimes end up with multiple processes signalling
walwriter, which can be bad, because it increases the likelihood that some of
the signals may be received when we are already holding WALWriteLock, delaying
its release...

That can only happen when walwriter has just come out of "hibernation", ie.
when the system has been idle for a while. So probably not a big deal in
practice.

Maybe I am missing something here - why can this only happen when hibernating?
Even outside of that two backends can decide that that they need to wake up
walwriter?

Ah sure, multiple backends can decide to wake up walwriter at the same
time. I thought you meant that the window for that was somehow wider
when XLogCtl->WalWriterSleeping.

We could prevent that, by updating state when requesting walwriter to be woken
up. But with the changes we're discussing below, that should be rare.

One small easy thing we could do to reduce the redundant wakeups: only
wake up walwriter if asyncXactLSN points to different page than
prevAsyncXactLSN.

--
Heikki Linnakangas
Neon (https://neon.tech)

On 25/10/2023 21:59, Andres Freund wrote:

See attached. It's the same logic as in your patch, just formulatd more
clearly IMHO.

Yep, makes sense!

Pushed this. Thanks for the investigation!

--
Heikki Linnakangas
Neon (https://neon.tech)

On 2023-11-27 17:55:34 +0200, Heikki Linnakangas wrote:

On 25/10/2023 21:59, Andres Freund wrote:

See attached. It's the same logic as in your patch, just formulatd more
clearly IMHO.

Yep, makes sense!

Pushed this. Thanks for the investigation!

Thanks!