Re: [PATCH] Revert default wal_sync_method to fdatasync on Linux 2.6.33+

Josh Berkus wrote:
> On 12/6/10 6:13 PM, Tom Lane wrote:
> > Josh Berkus <josh(at)agliodbs(dot)com> writes:
> >> OK, patch coming then. Right now test_fsync aborts when O_DIRECT fails.
> >> What should I have it do instead?
> >
> > Report that it fails, and keep testing the other methods.
>
> Patch attached. Includes a fair amount of comment cleanup, since
> existing comments did not meet our current project standards. Tests all
> 6 of the methods we support separately.
>
> Some questions, though:
>
> (1) Why are we doing the open_sync different-size write test? AFAIK,
> this doesn't match any behavior which PostgreSQL has.

I did that so we could see the impact of doing 2 8k writes that were
both fsync'ed vs doing one 16k write and then fsync:

Compare open_sync with different sizes:
open_sync 16k write 201.323/second
2 open_sync 8k writes 332.466/second

We often write multiple 8k WAL pages and then fsync on commit.

> (2) In this patch, I'm stepping down the number of loops which
> fsync_writethrough does by 90%. The reason for that was that on the
> platforms where I tested writethrough (desktop machines), doing 10,000
> loops took 15-20 *minutes*, which seems hard on the user. Would be easy
> to revert if you think it's a bad idea.
> Possibly auto-sizing the number of loops based on the first fsync test
> might be a good idea, but seems like going a bit too far.

Sure, I recently increased the number, probably too much.

> (3) Should the multi-descriptor test be using writethrough on platforms
> which support it?

Uh, I didn't think that would matter because the test is to test kernel
behavior of writing to one file descriptor and fsyncing using another.

--
Bruce Momjian <bruce(at)momjian(dot)us> http://momjian.us
EnterpriseDB http://enterprisedb.com

+ It's impossible for everything to be true. +