Managing scheduled routing with a high-level communications language

One of the most critical issues in multiprocessors today is managing the communications among processors. This thesis addresses the programming of systems (such as the NuMesh) that handle routing via high-speed reprogrammable scheduled routers on each node. A portable ‘communications language’, COP, used to express communications requirements, is presented. Using scheduled routers to control data motion through the network has two advantages. First, by performing offline path scheduling for messages based on information extracted from the application, latency and congestion can be improved. Second, by removing the need for online decision making, such routers can run at extremely high speeds, further decreasing latencies and increasing performance. Simple communication patterns known at compile time are relatively easy to schedule. However, data-dependent communications present more of a challenge. The compiler makes tradeoffs among different compilation techniques based on information in the communication program, including communication type and predicted traffic, as well as knowledge of the system size and layout. Further, the compiler chooses dynamically how to break up an application’s communications into sequential phases, using communication relationships expressed in the input language. The compiler generates code that efficiently handles changing phases and ensuring data integrity while doing so. This thesis shows that certain classes of applications, such as those that suffer from congestion with dynamic routers, show a significant decrease in run-time routing cycle count when using scheduled routing. Furthermore, reprogrammable scheduled routers are shown to provide a relatively general solution to application communication needs; applications with uncongested or data-dependent traffic are found to take approximately the same number of routing cycles, yielding a decrease in overall runtime given the potentially high cycle-speed of scheduled routers. Thesis Supervisor: Stephen A. Ward Title: Professor of Electrical Engineering and Computer Science