Summary: Performance of PPro
Dear Netters,
Many thanks for the numerous replies on my question on the performance of PPro.
In order of appearance:
Edgardo Garcia, Nico van Eikema Hommes, Pedro A M Vazquez, Thomas Huber, Thomas
Wagener, John M. McKelvey, Jochen Buehler, Joe Durant, Alan Hewat, Arne
Elofsson
Unfortunately, there are not very much "hard facts" I could summarize.
(A lot
of indirect comparisons and -in my eyes (forgive me)- more or less emotional
problems with LINUX or INTEL CPUs in general ....
In order not to fill the bandwidth I picked out some things:
****************************************************************************
On Jul 17, 10:31am, Pedro A M Vazquez wrote:
> Subject: [COMP-CHEM] CCL:G:performance of PPro (fwd)
> Hello
> We just started to benchmark a PPro200MHz for quantum chemistry
> calculations.
> While I've not completed all sets of tests I've results for the
> Stream benchmark (Memory bandwidth), Bonnie (I/O) and the 4 Gamess
> benchmarks.
> These results are for an Asustek Motherboard with 64M of RAM,
> an Adaptec 2940 SCSI2 adaptor and a NEC 2.1G SCSI2 hard disk running
> under FreeBSD2.1.0:
>
> Stream:(a g94 job running during the benchmark)
>
> Function Rate (MB/s) RMS time Min time Max time
> Assignment: 97.5238 0.1965 0.1641 0.2109
> Scaling : 85.3333 0.1969 0.1875 0.2031
> Summing : 87.7714 0.2813 0.2734 0.2969
> SAXPYing : 90.3529 0.2797 0.2656 0.2891
>
> Just for comparision, these are the results for an Alpha Server 1000 4/266
> DEC Unix3.2C: (idle machine)
>
> Function Rate (MB/s) RMS time Min time Max time
> Assignment: 96.0000 0.1801 0.1667 0.1833
> Scaling : 96.0000 0.1836 0.1667 0.2000
> Summing : 96.0000 0.2586 0.2500 0.2833
> SAXPYing : 96.0000 0.2618 0.2500 0.2667
>
> Bonnie:
> -------Sequential Output-------- ---Sequential Input--
--Random--
> -Per Char- --Block--- -Rewrite-- -Per Char- --Block---
--Seeks---
> Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec
%CPU
>
> 2940/PP200
> 100 3275 56.5 3194 12.8 1424 11.2 4978 80.4 4926 17.8 103.8
4.1
>
> Alpha 1000 4/266 DEC RZ28
> 100 4906 95.4 5439 17.4 2550 10.4 4746 82.6 5322 12.6 428.0
10.8
>
> GAMESS:
> Bench 10 04 13 07
> =============================================
> IBM370 64 74 366 391
> P166 61 78 281 335
> IBM580 59 69 322 381
>
> PPro200 43.6 69.3 163.7 254.6 <<<<<<
>
> IBM590 33 33 147 202
> Alpha1000 4/266 23.3 30.5 126.9 147.0
>
> The results for the Pentium166 (64M/RAM,FreeBSD2.1.0,2940/NEC2.1HD) and
> for the PPro200 were obtained with GAMESS (22/nov/95) compiled with
> GNU Fortran 0.5.18 (slightly faster than f2c/gcc).
>
> I need to benchmark gaussian yet to have a more representative sampling
> but as you can see these are very good results for a cheap computer.
> If you opt to buy 2 PPro to run GAMESS in parallel you'll got near
> two times the above reported performance.
>
> Pedro
>-- End of excerpt from Pedro A M Vazquez
On Jul 17, 3:57pm, Thomas Huber wrote:
> > > > Model: Plastocyanin all atom force-field with 3375 water
molecules
> > > > periodic boundary conditions, with a residue based
cutoff (1.2
nm)
> > > > resulting in an about 6,000,000 sized pairlist
> > > > 100 steps MD
> > > >
> > > > Machine Wallclock time [s]
> > > >
> > > > Pentium 90MHz
> > > > Plato board linux / f2c&gcc 2791
> > > > linux / g77 3235
> > > >
> > > > PentiumPro 150MHz
> > > > WinNT / Microsoft Fortran PowerStation 4.0 601
> > > >
> > > > HP 735 125MHz 539
> > > >
> > > > SGI R8000 75MHz 335
> > > >
> > > > Dec alpha 600 266MHz 378
> > > >
> > > > Cray C90 117
> > > >
> > > > Not too bad!!!!!
> > > >
> > > > Thomas
>-- End of excerpt from Thomas Huber
On Jul 17, 9:41am, Joe Durant wrote:
> Subject: Re: CCL:G:performance of PPro
> Hallo Rockus!
>
> We have been migrating to Pentiums and Pentium Pros... I put together
> a web-page with some of our benchmarks, look at
> http://mephisto.ca.sandia.gov/benchmarks.html
> I am really happy with my dual P6 box... each processor is roughly the
> speed of my R8000, and the whole thing costs about $6K. The dual P6
> boards offer 4 way interleaving on the memory, which I believe
> explains the speedup of the dual board over the single processor
> board. The loss in performance for things like QCISD is due, I am
> told, to the gcc compilers not yet being as good at optimizing matrix
> calculations. These same people expect the differences to disappear
> as the gnu compilers for the P6 mature.
>
> I look forward to your summary.
>
> Joe
>-- End of excerpt from Joe Durant
************************************************************************
Best greetings to all of you,
Rochus
--
********************************************************************************
Rochus Schmid
Technische Universitaet Muenchen Tel. ++49 89 2891 3140
Lehrstuhl f. Anorganische Chemie 1 Fax. ++49 89 2891 3088
Prof. W. A. Herrmann E-mail:
Lichtenbergstrasse 4 rochus (- at -) felix.anorg.chemie.tu-muenchen.de
85747 Garching
********************************************************************************