next up previous
Next: Embedded Software Applications Up: Introduction Previous: Block-Diagonal-Bordered Iterative Linear

Low-Latency Communications

For this work, active message-based communications on the Thinking Machines CM-5 provided implementations with extremely low-latency interprocessor communications. Separate algorithms have been developed for conventional non-blocking, buffered communications and low-latency, active message-based communications. Active message remote procedure calls (RPCs) provide protocol-less access to the transport layer of the CM-5 communication network. The user must assume responsibilities for all aspects of communications --- but is rewarded with very low-latency communications for short messages. The extremely low-latency communications available with active messages, permit a parallel-code development paradigm that greatly simplifies the implementation of these sparse matrix algorithms. Empirical data have been collected on both active message-based implementations and more traditional non-blocking, buffered message passing commands in order to illustrate the need for low-latency communications when solving matrices as sparse and irregular as power systems matrices.



David P. Koester
Sun Oct 22 17:27:14 EDT 1995