From: Osumi, Takamichi/大墨 昂道 <osumi.takamichi@fujitsu.com>
> I executed each wal_level three times and calculated the average time
> and found that disabling WAL logging reduced about 73 % of the minimal's
> loading speed
> in this test. This speed-up came from the difference of generated WAL sizes.
So, it's 4x speedup when the WAL buffer and/or disk is likely to be saturated. That's nice!
> In this test, to load the data generated more than 10GB of WALs with
> wal_level=minimal
> while wal_level=none emits just 760 bytes of WALs.
>
> I expect this size for none will become smaller when
> I take the modification to filter out the types of WAL which is discussed above.
I don't expect so, because the WAL volume is already very small in this workload (probably only the commit records.)
> Also, I monitored numbers of iostat's 'util' and noticed that
> util's spike to use I/O reduced from twice to once when I changed the level
> from minimal to none, which should be the effect of the patch.
It's a bit mysterious. I thought %util would be constantly 100% when wal_level = minimal. That may be because
wal_buffersis large.
Regards
Takayuki Tsunakawa