Re: [HACKERS] Performance while loading data and indexing

Justin Clift <justin@postgresql.org> writes:
>> On 26 Sep 2002 at 19:05, Martijn van Oosterhout wrote:
>>> fsync IIRC only affects the WAL buffers now but it may be quite expensive,
>>> especially considering it's running on every transaction commit. Oh, your
>>> WAL files are on a seperate disk from the data?

> Not sure if this is a good idea.  Would have to think deeply about the
> controller and drive optimisation/load characteristics.

> If it's any help, when I was testing recently with WAL on a separate
> drive, the WAL logs were doing more read&writes per second than the main
> data drive.

... but way fewer seeks.  For anything involving lots of updating
transactions (and certainly 5000 separate insertions per second would
qualify; can those be batched??), it should be a win to put WAL on its
own spindle, just to get locality of access to the WAL.

            regards, tom lane