> Actually, it seems to me that with the addition of the WAL in
PostgreSQL
> and the subsequent decreased need to fsync the data files themselves
> (only during checkpoints?), that the only time a battery-backed write
> cache would make a really large performance difference would be on the
> drive(s) hosting the WAL.
It still helps. In my experience a good BBU Raid controller is only
slightly slower than fsync=false. Fear the checkpoint storm if you
don't have some write caching. Beyond that I don't really care about
write delay.
Another thing to watch out for is that some sync modes (varying by
platform) can do >1 seeks per sync. This will absolutely kill your
commit performance on the WAL without write caching.
> So although it is in general good to have a dedicated spindle for the
> WAL, for many workloads it is in fact significantly better to have the
> WAL written to a battery-backed write cache. The exception would be
for
> applications with fewer, larger transactions, in which case you could
> actually use the dedicated spindle.
Exactly.
> Hmmm, on second thought, now I think I understand the rationale behind
> having a non-zero commit delay setting-- the problem with putting
I don't trust commit_delay. Get a good raid controller and make sure pg
is properly using it. Now, if you can't (or won't) do some type of
write caching bbu or no, your system has to be very carefully designed
to get any performance at all, especially with high transaction volumes.
Merlin
