Transcript Materials Computation Center R.M. Martin and J.P. Leburton

Materials Computation Center, University of Illinois
Duane Johnson and Richard Martin, NSF DMR-03-25939
Hijacking Game Consoles for Molecular Modeling
Todd J. Martinez (Chemistry) and S. Patel (ECE)
Research Objectives: Using game consoles as a
platform for molecular modeling. We have
concentrated on demonstrating the viability of the idea
using the Sony PlayStation2.
Approach: We have ported GAMESS, a well-known
open source quantum chemistry code to the PS2. We
have explored the use of hand-coded assembler for
linear algebra routines to make an efficient BLAS
package. Obstacles to overcome include the limited
options for fast data transfer due to the intended use
for gaming where the data direction is expected to be
exclusively from CPU to vector units (VU) to rendering
unit and VU to CPU transfer is not required.
Significant Results: We showed that performance
of up to 1GFLOP can be obtained for matrix-matrix
multiplication. We have shown that realistic
calculations can be carried out on the PS2. Parallel
performance with MPI was investigated and found to
be comparable to Linux PCs with similar network
connectivity.
Broader Impact: Low-cost, high-performance
platform for molecular modeling. One can expect that
this approach will lead to higher performance on even
a single processor because of the economics of game
consoles.
MCC website: http://www.mcc.uiuc.edu
Cluster of PS2’s running Linux and MPI
PS2
PIII600
P43200
Processor Clock (MHz)
300
600
3200
Floating Point Ops/Clock
20
2
2
MFLOPS
6000
1200
6400
System Bus Clock (MHz)
Bus Width (bits)
Data Transfer (MB/s)
150
128
2400
100
64
800
800
64
6400
32
256
1024
$500
$2,000
$2,000
Memory (MB)
Price (when first released)
©Board of Trustees University of Illinois
Materials Computation Center, University of Illinois
Duane Johnson and Richard Martin, NSF DMR-03-25939
Hijacking Game Consoles for Molecular Modeling
Todd J. Martinez (Chemistry) and S. Patel (ECE)
• Using a simulator, matrix-matrix multiplication
performance of more than 80% theoretical peak
was obtained using a single vector unit on the Cell
processor. Ongoing work at Yorktown Heights will
extend this code to cover all of the BLAS
operations.
Future Directions: The new vector-parallel
architecture calls for reworking of some of the
algorithms used, such as the construction of
electron-electron repulsion integrals. By continued
collaboration with IBM, we plan to be in a position to
use the PS3 for scientific computing as soon as
PS3’s are available for purchase. This is needed to
exploit the processor in a timely manner, before
traditional PCs can match its performance.
MCC website: http://www.mcc.uiuc.edu
PS2-GCC (17MFLOPS)
10
8
Time (s)
• We have begun a collaboration with IBM to
investigate the performance of the new Cell
architecture, which is planned for use in the nextgeneration game consoles (PlayStation3). Benjamin
Levine (UIUC) interned at IBM Yorktown Heights
with Jeff Burns and Ashwini Nanda.
6
PIII-600-BLAS (36MFLOPS)
4
2
PS2-VU0 (45-60 MFLOPS)
0
0
5
10
15
Millions of 4-element Dot Products
Dot Product Performance on PS2
GAMESS Running on a PS2 After Porting
©Board of Trustees University of Illinois
20