Significance of Results

Oracle's Sun Blade X6275 cluster is 28% faster than the IBM POWER6 cluster on Weather Research and Forecasting (WRF) continental United Status (CONUS) benchmark datasets.The Sun Blade X6275 cluster used a Quad Data Rate (QDR) InfiniBand connection along with Intel compilers and MPI.

  • On the 12 km CONUS data set, the Sun Blade X6275 cluster was28% faster than the IBM POWER6 cluster at 512 cores.

  • The Sun Blade X6275 cluster with 768 cores (one full Sun Blade6048 chassis) was 47% faster than 1024 cores of the IBMPOWER6 cluster (multiple racks).

  • On the 2.5 km CONUS data set, the Sun Blade X6275 cluster was21% faster than the IBM POWER6 cluster at 512 cores.

  • The Sun Blade X6275 cluster with 768 cores (on full Sun Blade6048 chassis) outperforms the IBM Power6 clusterwith 1024 cores by 28% on the 2.5 km CONUS dataset.

Performance Landscape

The performance in GFLOPS is shown below on multiple datasets.


Weather Research and Forecasting
CONUS 12 KM Dataset Cores Performance in GFLOPS Sun
X6275 Intel
Whitebox IBM
POWER6 Cray
XT5 SGI TACC
Ranger Blue
Gene/P 8 17.5 19.8 17.0
10.2

16 38.7 37.5 33.6 21.4 20.1 10.8
32 71.6 73.3 66.5 40.4 39.8 21.2 5.9 64 132.5 131.4 117.2 75.2 77.0 37.8
128 235.8 232.8 209.1 137.4 114.0 74.5 20.4 192 323.6





256 405.2 415.1 363.1 243.2 197.9 121.0 37.4 384 556.6





512 691.9 696.7 542.2 392.2 375.2 193.9 65.6 768 912.0






1024

618.5 634.1 605.9 271.7 108.5 1700



840.1


2048





175.6


All cores used on each node which participates in eachrun.

Sun X6275 - 2.93 GHz X5570, InfiniBand
Intel Whitebox - 2.8 GHz GHz X5560, InfiniBand
IBM POWER6 - IBM Power 575, 4.7 GHz POWER6, InfiniBand, 3 frames
Cray XT5 - 2.7 GHz AMD Opteron (Shanghai), Cray SeaStar 2.1
SGI - best of a variety of results
TACC Ranger - 2.3 GHz AMD Opteron (Barcelona), InfiniBand
Blue Gene/P - 850 MHz PowerPC 450, 3D-Torus (proprietary)


Weather Research and Forecasting
CONUS 2.5 KM Dataset Cores Performance in GFLOPS Sun
X6275 SGI
8200EX Blue
Gene/L IBM
POWER6 Cray
XT5 Intel
Whitebox TACC
Ranger 16 35.2






32 69.6

64.3


64 140.2

130.9
147.8 24.5 128 278.9 89.9
242.5 152.1 290.6 87.7 192 400.5





256 514.8 179.6 8.3 431.3 306.3 535.0 145.3 384 735.1





512 973.5 339.9 16.5 804.4 566.2 1019.9 311.0 768 1367.7





1024
721.5 124.8 1067.3 1075.9 1911.4 413.4 2048
1389.5 241.2
1849.7 3251.1
2600




4320.6
3072
1918.7 350.5
2651.3

4096
2543.5 453.2
3288.7

6144
3057.3 642.3
4280.1

8192
3569.7 820.4
5140.4

18432

1238.0



Sun X6275 - 2.93 GHz X5570, InfiniBand
SGI 8200EX - 3.0 GHz E5472, InfiniBand
Blue Gene/L - 700 MHz PowerPC 440, 3D-Torus (proprietary)
IBM POWER6 - IBM Power 575, 4.7 GHz POWER6, InfiniBand, 3 frames
Cray XT5 - 2.4 GHz AMD Opteron (Shanghai), Cray SeaStar 2.1
Intel Whitebox - 2.8 GHz GHz X5560, InfiniBand
TACC Ranger - 2.3 GHz AMD Opteron (Barcelona), InfiniBand


Results and Configuration Summary

Hardware Configuration:


48 x Sun Blade X6275 server modules, 2 nodes per blade, each node with

2 Intel Xeon X5570 2.93 GHz processors, turbo enabled, ht disabled
24 GB memory
QDR InfiniBand

Software Configuration:


SUSE Linux Enterprise Server 10 SP2
Intel Compilers 11.1.059
Intel MPI 3.2.2
WRF 3.0.1.1
WRF 3.1.1
netCDF 4.0.1

Benchmark Description

The Weather Research and Forecasting (WRF) Model is a next-generationmesoscale numerical weather prediction system designed to serve bothoperational forecasting and atmospheric research needs. WRF is designedto be a flexible, state-of-the-art atmospheric simulation system thatis portable and efficient on available parallel computing platforms. Itfeatures multiple dynamical cores, a 3-dimensional variational (3DVAR)data assimilation system, and a software architecture allowing forcomputational parallelism and system extensibility.

There are two fixed-size benchmark cases.



Single domain, medium size 12KM Continental US (CONUS-12K)

  • 425x300x35 cell volume
  • 48hr, 12km resolution dataset from Oct 24, 2001
  • Benchmark is a 3hr simulation for hrs 25-27 starting from a provided restart file
  • Iterations output at every 72 sec of simulation time, with the computation cost of each time step ~30 GFLOP

Single domain, large size 2.5KM Continental US (CONUS-2.5K)

  • 1501x1201x35 cell volume
  • 6hr, 2.5km resolution dataset from June 4, 2005
  • Benchmark is the final 3hr simulation for hrs 3-6 startingfrom a provided restart file; the benchmark may also beperformed (but seldom reported) for the full 6hrs startingfrom a cold start
  • Iterations output at every 15 sec of simulation time, withthe computation cost of each time step ~412 GFLOP
See Also

Disclosure Statement

WRF, see http://www.mmm.ucar.edu/wrf/WG2/bench/, resultsas of 3/8/2010.



Read More about [WRF Benchmark: X6275 Beats Power6...