Re: Itanium performance news...

From: Bill Todd (billtodd_at_metrocast.net)
Date: 07/31/03 

Date: Thu, 31 Jul 2003 03:39:55 -0400

"Mike Naime" <mnaime@kc.rr.com> wrote in message
news:Xh1Wa.46722$7O4.1106053@twister.rdc-kc.rr.com...
>
> JF Mezei <jfmezei.nospam@istop.com> wrote in message
> news:3F288452.1543611@istop.com...
> > re: "world record" done on some IA64 box.
> >
> > Can any objective person comment on whether Alpha EV7 could have beat
that
> > record had Digital still been a competitor of HP with a vested interest
in
> > marketing and making Alpha compete head on against intel ?%

Probably not without getting fired from HP, since while HP reportedly *has*
been doing some in-house TPC-C testing on Marvel systems with excellent
results they've announced that the results won't be made public.

One could look at at least somewhat-related benchmarks such as the 'Oracle
Applications Standard Benchmark' and see that, while an 8-processor EV7
systems only out-performs a 4-processor 1.5 GHz Itanic system by about 20%,
the lousy scaling of SuperDome on TPC-C from 4 to 64 processors would almost
certainly allow a 64-processor Marvel system (which I don't think is yet
available) to beat the pants off the new 64-processor SuperDome result - and
even a 32-processor Marvel system should have a shot at beating it.
Similarly, the fact that the 32-processor Marvel SAP SD result is over 5.23
times that of the 4-processor 1.5 GHz Itanic also suggests that
processor-for-processor Marvel would soundly trounce SuperDome (and should
at least come close to equaling its performance with only *half* as many
processors if SuperDome scales as poorly here as it does in TPC-C).

Of course, if Compaq hadn't dragged its feet from the time Curly took over
we'd most likely be comparing EV8's performance (in a 130 nm process, at 1.8
GHz) to Itanic's right now rather than EV7's (at 1.15 GHz in its 180 nm
process). If, that is, Itanic was still even afloat under those
circumstances.

> >
> > I don't trust HP to push EV7 to its limit since they woudln't want it to
> beat
> > Intel's chip.

Just because they apparently arbitrarily delayed EV7's introduction so as
not to overshadow McKinley's, and seemed uninterested in pushing EV7's clock
rate up to EV6's (let alone higher), and sacrificed EV8 to Satan Clara?
Naw...

>
> Well I'm not sure if this is comparing apples to apples.... But a
> comparison of newer AIX and VMS systems.
>
> Our peformance lab benchmarks IBM AIX and VMS systems with a simulated
> workload of our Application sending requests to the Oracle database. The
> test system is supposed to run for one hour under this simulated workload.
> This last spring, they did some testing of the IBM Reggata, and the
> Compaq/HP Marvel.
>
> The 16 CPU IBM Reggata system finished about 3/4 of the entire workload in
> one hour.
> The Marvel (GS1280) with less CPU's (I do not remember the number) was not
> rated at this time because it FINISHED the entire workload in about 45
> minutes! (Again I do not remember the exact timings, it may have been
less)
> It was determined that we would need another Marvel to feed the test
Marvel
> enough simulated workload in order to have enough work to keep running the
> test for the required one hour time.

If that was a 1.3 GHz Regatta system it's only fair to note that they're now
running at 1.7 GHz (with some other faster system components as well).
Still, it sounds as if at least on your application Marvel should handlily
out-perform the newest Regattas on a per-processor basis - and since those
new 32-processor Regattas are less than 8% slower on TPC-C than the new
'world record' 64-processor SuperDomes, if your application resembles TPC-C
then there's a good chance that a 32-processor Marvel system would beat that
new SuperDome 'world record'.

- bill