 |
|
|
|
1999 MCS Divisional Seminars & Colloquia |
|
Exploiting Shared Memory in Message-Passing ProgramsQuinn Snell 10:30 AM, Tuesday, February 9, 1999 |
|
| Abstract | Supercomputing has evolved from single shared memory machines or clusters of single processor machines to clusters of shared-memory multiprocessors. This evolution has yielded higher performance without drastically increasing cost or complexity. However, message passing programs and systems have not changed to exploit the inherent advantages of having shared memory on each node in the cluster. Current message passing is based solely on node number. Significant performance increases can be achieved by collecting messages from multiple processes on a single node and sending them collectively and also by using shared memory between processes on a node and only using the network for communication between nodes. In this talk, we propose an extension to MPI that exposes the shared- memory architecture to the programmer. The extension is also used to optimize collective communication routines. We present results using this approach that show as much as 300 percent performance increase. |
| [MCS | Research | Resources | People | Collaboration | Software | Publications | Information] | |||
|
