CFD Benchmark of 64-bit PC Cluster

1
CFD Benchmark of 64-bit PC Cluster
Yao-Huan Tseng Institute of Astronomy and
Astrophysics, Academia Sinica, ROC
Abstract Although the 64-bit x86 machines
have existed for few years, the price/performance
ratio is not acceptable for a small research
group like us and taking advantage of 64-bit x86
machines is not in a reasonable way until AMD
64-bit CPU present. The AMD 64-bit Opteron is a
new device family based on a new 64-bit
architecture that is compatible with the
pre-existing x86 32-bit architecture. The most
exciting feature of the Opteron is the build-in
memory controller which is totally different
strategy from Intel cpu and allows to maximize
the performance of SMP machines. Here we present
results from a few benchmark tests for running
high order Godunov CFD codes and parallel VH-1
codes on a variety of 32/64-bit PCs.

• Hardware
• 8 single-cpu cluster (2.6GHz P4, 2GB DDR,
  2x120GB IDE, Gbps NIC)
• 10 dual-cpu 32bits cluster (2.4GHz Xeon, 2GB
  DDR, 120GB IDE, Gbps NIC)
• 5 dual-cpu 64bits cluster (1.8GHz AMD Opetron,
  2GB DDR, 120GB IDE, Gbps NIC)
• 24 nodes Linpack Benchmark (60.51Gflops)

Single CPU benchmark A High order Godunov
method 250x180 resolution B Relax method
250x180 resolution C High order Godunov method
1024x1024 resolution D Relax method
1024x1024 resolution
A
B
C
D
INTEL
INTEL
INTEL
INTEL
AMD
AMD
AMD
AMD
Dual CPU benchmark A High order Godunov method
250 x 180 resolution B Relax method 250 x
180 resolution C High order Godunov method
1024 x 1024 resolution D Relax method 1024 x
1024 resolution
A
B
C
D
Parallel benchmark A High order Godunov method
512 x 512 B VH1(PPM) 1920 x 1920
INTEL
INTEL
AMD
AMD

• Summary
• Although the difference frequency of CPUs, the
  single-cpu performance of AMD/Opteron(1.8GHz) is
  about ? of single Intel Xeon(2.4GHz)
• In dual-cpu test, two intel Xeon take ? 1.6
  execution time of one Xeon for larger scale
  problem. But almost no difference of execution
  time between AMD/Opteron running two jobs at the
  same time and running one job.
• The performance scaling of the two clusters are
  almost the same for more CPUs involved. The
  overall performance of AMD/Opteron is superior to
  Intel/Xeon, not only the performance of single
  CPU but also the Opteron's NUMA memory
  architecture.

2004/11/30