Hiding Communication Latency With Non-SPMD, Graph-Based Execution
Source: University of California
Reformulating an algorithm to mask communication delays is crucial in maintaining scalability, but traditional solutions embed the overlap strategy into the application. The authors present an alternative approach based on data flow, which factors the overlap strategy out of the application. Using this approach they are able to reduce communication delays, meeting and in many cases exceeding performance obtained with traditional hand coded applications. Spurred on by the multi-core processor, scalable systems have the potential to enable simulations of remarkable fidelity and complexity, leading to new scientific discovery. However, improvements in processor performance amplify the cost of o -chip data motion, and applications must cope by with this trend by tolerating latency.