Dear Szabolcs,
There is an evaluation note that Hans Wenzel did for US-CMS.
It is a CMS Internal note. It's a megabyte so I won't forward it to
the list.
http://cmsdoc.cern.ch/documents/05/in/in05_012.pdf
As a result of these evaluations, US-CMS decided to buy 280 Dual
Opteron nodes. The gatekeepers are still on Xeon systems. These
are used for worker nodes. The grid job manager was modified to
submit to either architecture. So far the experience is fine, the
applications run transparently. We have put only one rack in
production. With three more racks arriving later this week. We
don't have enough experience to judge reliability but so far there
are no indications of problems.
We did evaluations of power, but they didn't become an internal note.
At the bottom are the most recent results that Hans did with Dual
Core Opterons, which US-CMS has purchased at one of the Tier-2
centers. The initial indications of performance and power
consumption are very promising.
-Ian
Root (64/64 bit)
ROOTMARKS
single bench 1157.83
single geom 1332.4
single linear 1141.6
single stress 1201.6
# stress (simul) Rootmarks:
eff (compared to single process)
1
1201.6 100 %
2
1216.9/1186.6 100.01 %
3
1186.6/1199.2/1187.0 99.11 %
4
1190.4/1179.8/1183.9/1177.2 98.44 %
# pythia (simul) Evts/sec
eff (compared to single process)
1
122.95 100%
2
122.9/123 100%
3
123.16/122.97/122.95 100%
4
123.03/123.07/122.96/123.21 100%
# oscar (simul) time (sec) evts/
sec efficiency in % (compared to single proc)
------------------------------------------------------------------------
----------------
1 2952.20 0.1016 evts/
sec 100 %
2 2959.96 0.1014 evts/
sec 99.6 %
2 2971.21 0.1010 evts/sec
3 2993.83 0.1002 evts/sec 98.29 %
3 3016.94 0.0994 evts/sec
3 2999.68 0.1000 evts/sec
4 3030.53 0.0990 evts/sec 97.8 %
4 3029.41 0.0990 evts/sec
4 3017.96 0.0994 evts/sec
4 2997.97 0.1001 evts/sec
On Aug 25, 2005, at 10:13 AM, Hernath Szabolcs wrote:
> Dear List,
>
> we are about to spend a lot of money on LCG resources, and I would
> like to hear your comments about our choice:
>
> 64 bit AMD (Opteron) vs. Intel Xeon.
>
> 1. SLC/LCG middleware:
>
> As far as I know, Opterons are 100% compatible with 32bit
> applications, but it isn't completely clear to me whether SLC/LCG
> middleware can be run right now in 64 bit mode. If not, does anyone
> have an idea whether it is likely to be supported in the near future?
>
> 2. Performance:
>
> Any real-world experience on the performance of AMD64/32 bit mode
> vs. AMD64/64 bit mode vs. Intel Xeon is welcome.
>
> 3. Power consumption/heat dissipation:
>
> see point 2. ;-)
>
> 4. Reliability, other issues etc:
> ...
>
> 5. Any arguments for IA64?
>
> Thanks in advance,
> Cheers
>
> Szabolcs
>
>
|