Mark Rae wrote:
> On Tue, Mar 15, 2005 at 12:51:03PM -0800, Jeff Davis wrote:
> > Be careful assuming that. DB benchmarks are hard to do in a general
> > sense. His results probably indicate a general trend, but you should
> > test your application yourself to get a real result. His pattern of SQL
> > queries might be very different from yours.
>
> Very true.
>
> You may have noticed that I had a very low query rate of 5.8 queries
> per second, because some of the queries have 12 tables to join and
> take about 20s to run. This tends to work in postgres' favour.
> If you have lots have simple queries, it will be better for mysql
> and the break even point will be higher.
>
>
> Also, while on the subject of scaling. I had the opportunity
> to try postgres on a 16CPU Altix and couldn't get it to scale
> more than about 4x, whereas Oracle got up to about 12x faster
>
> I assume this is because of the NUMA architecture. I was also
> told that Oracle had made no special optimizations to accomodate it.
>
> My guess is that because postgres allocates all its shared
> buffers as a contiguous chunk, it puts all the load on one
> memory bank.
> Oracle on the other hand, seems to use lots of smaller regions
> which would probably be spread throughout the physical memory.
>
> Perhaps one of the developers could comment on how difficult
> it would be to change the shared buffer handling to use multiple
> segments. As I'd definitely be willing to give it a go.
We have had some major SMP improvements in current CVS. Were you
testing that or 8.0.X?
--
Bruce Momjian | http://candle.pha.pa.us
pgman@candle.pha.pa.us | (610) 359-1001
+ If your life is a hard drive, | 13 Roberts Road
+ Christ can be your backup. | Newtown Square, Pennsylvania 19073
