Jeff Davis <pgsql@j-davis.com> writes:
> For the case of initdb, the data needing to be fsync'd is effectively
> constant, and it's a lot of small files. If the requests don't make it
> to the io scheduler queue before fsync, the kernel doesn't have an
> opportunity to schedule them properly.
> But for larger amounts of data copying (like ALTER DATABASE SET
> TABLESPACE), it seemed like there was more risk that sync_file_range
> would starve out other processes by continuously filling up the io
> scheduler queue (I'm not sure if there are protections against that or
> not). Also, if the files are larger, a single fsync represents more
> data, and the kernel may be able to schedule it well enough anyway.
> I'm not an authority in this area though, so if you are comfortable
> extending sync_file_range to copydir() as well, that's fine with me.
It could use some performance testing, which I don't have the time
for (or proper equipment). Anyone?
Also, I note that copy_file is set up so that it does sync_file_range or
fadvise for each 64K chunk of data, which seems mighty small. I wonder
if it'd be better to take that out of the loop and do one whole-file
advise at the end of the copy loop. Or at least use some larger
granularity for those calls.
regards, tom lane