Re: pgbench Windows outperforms Debian in Write tests while Debian is faster in SELECT only tests

"Klaus P. Pieper" <kpi6288(at)gmail(dot)com> writes:
>> Are you sure you have Windows configured to fsync properly?

> It seems that it is indeed related to the wal_sync_method settings.

> Results for 1 client read / write:

> Option Windows Debian
> fdatasync 340 150
> fsync 125 90
> open_datasync 360 160
> fsync_writethrough 115 n/a
> open_sync n/a 90
> fsync = off 360 400

OK, so open_datasync and fdatasync are nonfunctional on Windows, which
is unsurprising. The other cases are evidently doing *something*, but
the amount of overhead is variable from one sync method to another,
which is likewise unsurprising. The whole reason why we have these
options at all is that some systems handle some of them better than
others. In this case, I think the relevant conclusion is that correct
configuration for Debian is to use open_datasync or fdatasync, while
correct configuration for Windows is to use fsync or fsync_writethrough,
and when you compare those two configurations, yes Windows is slower.

If you're a paranoid sort you would want to do plug-pull testing
to ensure that your selected setting actually does prevent data
corruption during a system crash.

regards, tom lane