BNL CASPUR CERN PricePerformance estimates for some compute platforms AMD, Intel

1
BNL / CASPUR / CERNPrice/Performance estimates
for some compute platforms (AMD, Intel)

• Pavel Nevski
• BNL /CERN
• October 2004

2
Participated BNL P.Nevski CASPUR
A.Maslennikov(), M.Rosati CERN
E.McIntosh, I.McLaren ()
Project Coordinator
3
Goals
We have tried to compare the two most
interesting platforms of the day Intel Nocona
and AMD Opteron In short, we wanted to check
which of the two may deliver more compute power
per dollar.
4
Hardware under test (CASPUR)
Intel Nocona 2 CPUs at 3.4 GHz, 1GB RAM
RedHat
Enterprise 3
Kernel 2.4.21
Estimated price () 3200 Euro
(1 GB
RAM, incl. IPMI card, rack-mounted) AMD
Opteron 2 CPUs at 2 GHz, 1GB RAM
SuSE Enterprise
Edition 9.0
Kernel 2.6.5-7
SuSE numactl
Estimated price () 3200
Euro
(1 GB RAM, incl. IPMI function, rack-mounted)
() Prices as calculated by E4 Computer
Engineering, Italy
5
Benchmark suites
ATLSIM A full-scale GEANT3 simulation of
ATLAS detector (P.Nevski)
(typical LHC Higgs events) SixTrack
Tracking of two particles in a 6-dimensional
phase space including
synchrotron oscillations (F.Schmidt)
(http//frs.home.cern.ch/frs/sixtrac
k.html) Sixtrack
benchmark code E.McIntosh
(http//frs.home.cern.ch/frs/Benchmark/benchma
rk.html) CERN U Ancient CERN Units
Benchmark (E.McIntosh)
6
What was measured
On both platforms, we were running one or two
simultaneous jobs for each of the benchmarks.
On Opteron, we used the SuSE numactl
interface to make sure that at any time each of
the two processors makes use of the right bank of
memory. Example of submission, 2 simultaneous
jobs Intel ./TestJob ./TestJob AMD numactl
cpubind0 membind0 ./TestJob
numactl cpubind1 membind1 ./TestJob

7
Results
While both machines behave in a similar way
when only one job is run, the situation changes
in a visible manner in the case of two jobs. It
may take up to 30 more time to run two
simultaneous jobs on Intel, while on AMD there is
a notable absence of any visible performance
drop.
8
Conclusions
The results we have obtained were predictable.
The Intel platform is a plain SMP architecture,
and we are observing an apparent contention
effect on the memory bus and/or their single
memory controller, when running two simultaneous
tasks. The new AMD platform implements a
scalable NUMA architecture and allows for
independent and hence non-interfering
CPU-to-memory data exchanges over a pair of
memory controllers and Hypertransport buses.
As both machines under test cost practically
the same, we conclude that going Opteron today
may save us up to 30 for the same amount of
Teraflops.