Classification and Performance Evaluation of Hybrid Dataflow Techniques With Respect to Matrix Multiplication

This paper classifies hybrid dataflow techniques due to the instruction issuing technique. A software simulation is conducted to compare fine-grain dataflow to several hybrid dataflow techniques: multithreaded dataflow with direct token recycling as used in Monsoon, multithreaded dataflow with consecutive execution of the instructions within a thread as used in the Epsilon processors and in EM-4, dataflow with complex machine operations as proposed for the SIGMA-1, and large-grain dataflow presuming a RISC processor respectively a superscalar processor in the execution stage. All dataflow techniques show good scalability and effectively compensate delays caused by the network and by structure access, provided that load is sufficient. The achieved performance accelerations differ as follows: Large-grain dataflow proved superior to all other techniques provided that a superscalar processor is used, and performs at least equal to the other techniques with a RISC processor. Multithreaded dataflow is an improvement over fine-grain dataflow but suffers from a large overhead and thus does not achieve a considerable speedup. This is even worse for the cycle-by-cycle interleaving technique of Monsoon. Complex machine operations perform slightly better than large-grain dataflow with a RISC processor. It can be an useful enhancement to other techniques.