MILC
Walltime |
Speedup |
Relative Speed* |
*Speed per processor normalized to Datastar.
“The MILC Code is a body of high performance research software written in C for doing SU(3) lattice gauge theory on several different (MIMD) parallel computers in current use.” [MILC]
MILC Version 6 was used for the benchmarks. To optimize performance, a special library of assembly-coded matrix and vector routines is available for some types of processors. These routines were used on DataStar, while C versions of the routines were used on the other computers.
The two problems considered are from the NERSC benchmark set. The medium problem uses a 32^4 lattice, while the large problem uses a 64^4 lattice.
The table lists run times on the various computers for both problem sizes. The plots show the same results converted to relative speeds per processor in strong scaling scans.
Cobalt is the fastest on the medium problem, while T2 is fastest for the large problem. The communication fraction is modest for the processor counts considered, so all of the computers exhibit good scalability on both problems. Indeed, most of the computers show superlinear speedup for the medium problem, presumably because of better cache utilization with increasing processor count.
[MILC] The MILC Code (version: 6.20 sep02),
- http://www.physics.utah.edu/~detar/milc/milcv6.html
