| |
|
|
|
|
 |
NCSA NEWS |
|
|
|
Test runs on the Origin with two differing protocols for interprocessor communication (MPI and SHMEM) show a 60 percent speedup over the T3E with SHMEM protocol. The MPI version, which is more robust on the Origin for nondedicated modes, performs comparably to the T3E.
Although the T3E architecture is more efficient for interprocessor communication, says Beris, the Origin outdoes it with higher performance per processor, due to larger cache-size. "In order to achieve high performance," explains Beris, "our code requires extensive use of the cache. We have many vectorizable operations that can be greatly improved by the larger cache, whereas on the T3E, due to the smaller cache, this wasn't possible, and we had lower single-processor performance." Beris expects further improvement with code optimization.
|
|
|
|
|
|
