Re: Server recommendations

On Mon, 2003-10-06 at 08:32, Shridhar Daithankar wrote:
> Ron Johnson wrote:
> > On Mon, 2003-10-06 at 01:43, Shridhar Daithankar wrote:
>
> >>The best performance is by running 32 bit applications on 64 bit kernel/hardware
> >>, according to a migration guide by HP. The reasoning is using space optimally
> > Does HP have any AMD64 servers?
>
> No. But while migrating from PA1.1 to PA2.0, they have had that pain..:-) And I
> just went thr. opteron optimization guide from a link you provided. Basically
> same stuff HP was pitching.
>
> And I also configured a system on laclinux.com, just for curiosity. 2x
> 1.8GHz/4GB/36GBx2 10K RPM SCSI/Adaptec 29320 for $4800/- is damn cheap a system.
> HP does not even start below $5K for 64 bit systems.
>
> >>Imagine, if every long in pg is 8byte that would be waste most of the times.
> >>However given a native 8 byte integer/float is available, there is no reason to
> >>use a 8 byte data type unless required.
> >>From what I've read, longs are still 32-bit; it's only pointers
> > that have upped to 64-bit.
>
> I thought under 64 bits short==2 bytes, int==4 bytes and long==8 bytes. i.e.
> this is another architecture/word length where int != long in size.

Sounds like we are saying the same thing.

> >>Its about exploiting wide and fast bus of a 64bit machine in a most optimal
> >>fashion. I think except for kernel and glibc, nothing else requires 64 bit in
> >>general unless application insists on doing it's own caching.
> > In PG's case, if the app uses BIGINT a lot, then 64-bit PG should
> > be more efficient.
>
> Possibly. Source code availability has a great advantage here. Compile the way
> you want.
>
> But still, as long as PG in itself does not use 8 byte data types exclusively, I
> would like it to be a 32 bit app. A bigint is 8 byte on a 64 bit machine,
> whether or not the app. is 32 bit.
>
> > Besides, in 64-bit mode, the compilers get to use 2x as many GP
> > registers, which should increase performance.
>
> Well, if a 32 bit app. is optimised for opteron, compiler will use those
> registers anyway..:-) At least on HP, such trick can be done. IMO thats very
> handy..

Don't think it works that way in AMD64. The way I've read it, you
only get "32-bit apps" when you compile with the 32-bit compiler
switches, and thus it doesn't use the extra GPRs.

--
-----------------------------------------------------------------
Ron Johnson, Jr. ron(dot)l(dot)johnson(at)cox(dot)net
Jefferson, LA USA

"Fair is where you take your cows to be judged."
Unknown