Jim C. Nasby wrote:
I think you mean this...
http://www.postgresql.org/docs/8.1/static/runtime-config-wal.html
commit_delay (integer)
No, that's not what I mean at all. On a system doing a large number of
WAL-generating transactions per second, it's certainly possible for
multiple transactions to commit in the period of time it takes for the
platter to rotate back into position to allow for writing of the WAL
data. What I don't know is if those multiple transactions would actually
make it to the platter on that rotation, or if they'd serialize,
resulting in one commit per revolution. I do know that there's no
theoretical reason that they couldn't, it's just a matter of putting
enough intelligence in the drive.
Perhaps this is something that SCSI supports and (S)ATA doesn't, since
SCSI allows multiple transactions to be 'in flight' on the bus at once.
SCSI Command queueing:
http://www.storagereview.com/guide2000/ref/hdd/if/scsi/protCQR.html
SATA "native command queuing":
http://www.tomshardware.com/2004/11/16/can_command_queuing_turbo_charge_sata/
But since you mention commit_delay, this does lead to an interesting
possible use: set it equal to the effective rotational period of the
drive. If you know your transaction load well enough, you could possibly
gain some benefit here. But of course a RAID controller with a BBU would
be a better bet...
I suppose as long as you always have several transactions trying to
commit, have a separate spindle(s) for the WAL then you could improve
throughput at the cost of the shortest transaction times. Of course, it
might be that the increase in lock duration etc. might outweigh any
benefits. I'd suspect the cost/gain would be highly variable with
changes in workload, and as you say write-cache+BBU seems more sensible.
--
Richard Huxton
Archonet Ltd
