Обсуждение: Using condition variables to wait for checkpoints

Hello hackers,

A user complained about CREATE DATABASE taking > 200ms even with fsync
set to off.  Andres pointed out that that'd be the clunky poll/sleep
loops in checkpointer.c.

Here's a draft patch to use condition variables instead.

Unpatched:

postgres=# checkpoint;
CHECKPOINT
Time: 101.848 ms

Patched:

postgres=# checkpoint;
CHECKPOINT
Time: 1.851 ms

-- 
Thomas Munro
https://enterprisedb.com

Hi,

On 2019-03-13 11:56:19 +1300, Thomas Munro wrote:
> A user complained about CREATE DATABASE taking > 200ms even with fsync
> set to off.  Andres pointed out that that'd be the clunky poll/sleep
> loops in checkpointer.c.
> 
> Here's a draft patch to use condition variables instead.
> 
> Unpatched:
> 
> postgres=# checkpoint;
> CHECKPOINT
> Time: 101.848 ms
> 
> Patched:
> 
> postgres=# checkpoint;
> CHECKPOINT
> Time: 1.851 ms

Neat. That's with tiny shmem though, I bet?


> +        <row>
> +         <entry><literal>CheckpointDone</literal></entry>
> +         <entry>Waiting for a checkpoint to complete.</entry>
> +        </row>

> +        <row>
> +         <entry><literal>CheckpointStart</literal></entry>
> +         <entry>Waiting for a checkpoint to start.</entry>
> +        </row>

Not sure I like these much, but I can't quite ome up with something
meaningfully better.


Looks good to me. Having useful infrastructure is sure cool.


- Andres

On Tue, Mar 12, 2019 at 7:12 PM Andres Freund <andres@anarazel.de> wrote:
> Having useful infrastructure is sure cool.

Yay!

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Thu, Mar 14, 2019 at 1:15 AM Robert Haas <robertmhaas@gmail.com> wrote:
> On Tue, Mar 12, 2019 at 7:12 PM Andres Freund <andres@anarazel.de> wrote:
> > Having useful infrastructure is sure cool.
>
> Yay!

+1

I renamed the CVs because the names I had used before broke the
convention that variables named ckpt_* are protected by ckpt_lck, and
pushed.

There are some other things like this in the tree (grepping for
poll/pg_usleep loops finds examples in xlog.c, standby.c, ...).  That
might be worth looking into.

-- 
Thomas Munro
https://enterprisedb.com

On Thu, Mar 14, 2019 at 11:05 AM Thomas Munro <thomas.munro@gmail.com> wrote:
> I renamed the CVs because the names I had used before broke the
> convention that variables named ckpt_* are protected by ckpt_lck, and
> pushed.

Erm... this made successful checkpoints slightly faster but failed
checkpoints infinitely slower.  It would help if we woke up CV waiters
in the error path too.  Patch attached.

-- 
Thomas Munro
https://enterprisedb.com