Using opcode information to control thread scheduling in pipelined multithreaded processors

This paper examines the performance of two Delay Enforced MUltiStream (DEMUS) thread scheduling mechanisms for pipelined multithreaded processors. The Fixed-Delay (FD) mechanism, which is equivalent to the pessimistic thread scheduling technique of classical cyclic pipeline computers, is simple to implement, but requires as many running threads as pipeline stages for complete pipeline utilization. It is demonstrated by means of simulation that even the simple technique of scheduling based on opcode, as in the Modified Fixed-Delay (MFD) mechanism, is sufficient to greatly improve the performance characteristics of a pipelined multithreaded machine.