Compile time vs. runtime: scheduling parallelism on dataflow machines

Summary form only given. Exploitation of parallelism in programs decreases execution time but increases resource requirements. It is argued that this problem should be addressed by applying compile-time analysis and static scheduling techniques. Such techniques have been developed for scheduling loops on very long instruction word (VLIW) machines which, like dataflow machines, can exploit fine-grain parallelism in programs. In VLIW machines, all resources can be scheduled very precisely because of the synchronous model of parallelism. It is shown how these techniques can be applied to dataflow machines by adopting a less rigid notion of scheduling providing a reasonable trade-off between parallelism and resource usage. The authors also show how such asynchronous scheduling can be implemented in a novel dataflow loop schema for the Monsoon Explicit Token Store dataflow machine