Tom Lane kirjutas P, 13.04.2003 kell 18:45:
> [ Warning, topic drift ahead ]
>
> Shridhar Daithankar <shridhar_daithankar@persistent.co.in> writes:
> > However this would not work in all cases unless you are able to partition the
> > data. Otherwise you need a database that can have single database image
> > across machines.
>
> > If and when postgresql moves to mmap based model, postgresql running on mosix
> > should be able to do it.
>
> In a thread that's been criticizing handwavy arguments for fundamental
> redesigns offering dubious performance improvements, you should know
> better than to say such a thing ;-)
>
> I don't believe that such a design would work at all, much less have
> any confidence that it would give acceptable performance. Would mosix
> shared memory support TAS mutexes? I don't see how it could, really.
> That leaves you needing to come up with some other low-level lock
> mechanism and get it to have adequate performance across CPUs.
Does anybody have any idea how Oracle RAC does it ?
They seem to need to syncronize a lot (at least locks and data cache
coherency) across different machines.
> Even
> after you've got the locking to work, what would performance be like?
> Postgres is built on the assumption of cheap access to shared data
> structures (lock manager, buffer manager, etc) and I don't think this'll
> qualify as cheap.
[OT]
I vaguely remember some messages about getting PG to work well on NUMA
computers, which by definition should have non-uniformly cheap access to
shared data structures.
They must have faced similar problems.
-------------
Hannu