Project: Adaptive Data Placement
Investigators: David Lowenthal and Greg Andrews

The problem of data placement is vital to current parallel machines. A good data placement minimizes completion time by attempting to simultaneously minimize communication and balance the computational load.

There are generally two current approaches to data placement. Some languages provide user annotations to inform the compiler of the data placement, and some languages rely completely on the compiler to infer a good data placement. Both approaches have their drawbacks.

We are currently investigating a new adaptive approach. It is implemented in the Adapt system, which finds the best data placement at run time. When necessary, Adapt automatically changes the data placement in mid-application, so that it adapts to the needs of the application. Initial results show our adaptive approach to be superior to static approaches for applications requiring run-time support and competitive with static approaches for applications amenable to static analysis. Further details can be found in the following paper:

David K. Lowenthal and Gregory R. Andrews. Adaptive Data Placement for Distributed-Memory Machines. To appear, 10th International Parallel Processing Symposium, Honolulu, Hawaii, April 15-19, 1996 (192K bytes, 77K compressed)