Обсуждение: performance regression with Linux 2.6.33 and glibc 2.12

Hi.

I hope I'm not going to expose an already known problem, but I couldn't find
it mailing list archives (I only found http://archives.postgresql.org/pgsql-
hackers/2009-12/msg01543.php).

On one of my (non production) machines, I've just seen a very big performance
regression (I was doing a very simple insert test). I had an 'old' 8.4
postgresql compiled a few month ago, performing very well, and my 'bleeding
edge' 9.0, doing the same insert very slowly.

I managed to find the cause of the regression : with Linux 2.6.33, O_DSYNC is
now available. With glibc 2.12, O_DSYNC is available in userspace. Having both
(they are both very new, 2.12 isn't even official on glibc website), my new
build defaulted to open_datasync. The problem is that it is much slower. I
tested it on 2 small machines (no big raid, just basic machines, with SATA or
software RAID).

Here is the trivial test :
The configuration is the default configuration, just after initdb

CREATE TABLE test (a int);
CREATE INDEX idxtest on test (a);



with wal_sync_method = open_datasync (new default)

marc=# INSERT INTO test SELECT generate_series(1,100000);
INSERT 0 100000
Time: 16083,912 ms

with wal_sync_method = fdatasync (old default)

marc=# INSERT INTO test SELECT generate_series(1,100000);
INSERT 0 100000
Time: 954,000 ms

Doing synthetic benchmarks with test_fsync:

open_datasync performance, glibc 2.12, 2.6.34, 1 SATA drive

Simple 8k write timing:
write                           0.037511

Compare file sync methods using one 8k write:
open_datasync write            56.998797
open_sync write               168.653995
write, fdatasync               55.359279
write, fsync                  166.854911

Compare file sync methods using two 8k writes:
open_datasync write, write    113.342738
open_sync write, write        339.066883
write, write, fdatasync        57.336820
write, write, fsync           166.847923

Compare open_sync sizes:
                                                                       
16k open_sync write           169.423723
                                                               
2 8k open_sync writes         336.457119
                                                               

Compare fsync times on write() and new file descriptors (if the times
are similar, fsync() can sync data written on a different descriptor):
write, fsync, close           166.264048
write, close, fsync           168.702035

This is it, I just wanted to raise an alert on this: the degradation was 16-
fold with this test. We wont see linux 2.6.33 + glibc 2.12 in production
before months (I hope), but shouldn't PostgreSQL use fdatasync by default with
Linux, seeing the result ?

By the way, I re-did my tests with both 2.6.33, 2.6.34 and 2.6.35-rc1 and got
the exact same result (O_DSYNC there, obviously, but also the performance
degradation).

Cheers

Marc

Marc Cousin <cousinmarc@gmail.com> writes:
> I hope I'm not going to expose an already known problem, but I couldn't find
> it mailing list archives (I only found http://archives.postgresql.org/pgsql-
> hackers/2009-12/msg01543.php).

You sure this isn't the well-known "ext4 actually implements fsync
where ext3 didn't" issue?

            regards, tom lane

On Friday 04 June 2010 15:59:05 Tom Lane wrote:
> Marc Cousin <cousinmarc@gmail.com> writes:
> > I hope I'm not going to expose an already known problem, but I couldn't
> > find it mailing list archives (I only found
> > http://archives.postgresql.org/pgsql- hackers/2009-12/msg01543.php).
>
> You sure this isn't the well-known "ext4 actually implements fsync
> where ext3 didn't" issue?
I doubt it. It reads to me like he is testing the two methods on the same
installation with the same kernel

> > with wal_sync_method = open_datasync (new default)
> > marc=# INSERT INTO test SELECT generate_series(1,100000);
> > INSERT 0 100000
> > Time: 16083,912 ms
> >
> > with wal_sync_method = fdatasync (old default)
> >
> > marc=# INSERT INTO test SELECT generate_series(1,100000);
> > INSERT 0 100000
> > Time: 954,000 ms
Its not actually surprising that in such a open_datasync is hugely slower than
fdatasync. With open_datasync every single write will be synchronous, very
likely not reordered/batched/whatever. In contrast to that with fdatasync it
will only synced in way much bigger batches.

Or am I missing something?

I always thought the synchronous write methods to be a fallback kludge and
didnt realize its actually the preferred method...

Andres

The Friday 04 June 2010 15:59:05, Tom Lane wrote :
> Marc Cousin <cousinmarc@gmail.com> writes:
> > I hope I'm not going to expose an already known problem, but I couldn't
> > find it mailing list archives (I only found
> > http://archives.postgresql.org/pgsql- hackers/2009-12/msg01543.php).
>
> You sure this isn't the well-known "ext4 actually implements fsync
> where ext3 didn't" issue?
>
>             regards, tom lane

Everything is ext4. So I should have fsync working with write barriers on all
the tests.

I don't think this problem is of the same kind: I think it is really because
of O_DSYNC appearing on 2.6.33, and PostgreSQL using it by default now. If my
filesystem was lying to me about barriers, I should take no more performance
hit with open_datasync than with fdatasync, should I ?