On Mon, 19 May 2003, george young wrote:
> Has anyone run postgres on a beowulf system?
>
> I'm shopping for a new server. One candidate would be a
> quad opteron (64-bit AMD "hammer") machine. Another approach might
> be a beowulf of single or dual opterons. I imagine the beowulf
> would be a bit cheaper, and much more expandable, but what about
> the shared memory used by the postgres backends? I gather that
> postgres uses shared memory to coordinate (locks?) between backends?
>
> I have a smallish DB (pgdump|bzip2 -> 10MB), with ~45 users logged in
> using local X(python/gtk) postgres client apps.
>
> Will the much slower shared memory access between beowulf nodes be
> a performance bottleneck?
Save yourself some money on the big boxes and get a fast drive subsystem
and lots of memory, those are more important than raw horsepower, and any
dual Opteron / Itanium2 / USparc III / PPC / Xeon machine has plenty of
CPU ponies to handle the load.
We use dual PIII's for most of our serving, and while our front end web
servers need to grow a bit to handle all the PHP we're throwing at them,
the postgresql database on the dual PIII-750 is still plenty fast. I.e.
our bottlenecks are elsewhere than pgsql.
I don't know anyone off the top of my head that's running postgresql on an
Opteron, by the way, but I expect it should work fine. You're more likely
to have problems finding a distribution that works well on top of an
Opteron than to have problems with pgsql.