Speculative multithreaded architecture (SpMT) philosophy relies on aggressive speculative execution for improved performance. Aggressive speculative execution results in a significant wastage of dynamic energy due to useless computation in the event of mis-speculation. As energy consumption is becoming an important constraint in microprocessor design, it is extremely important to reduce such wastage of dynamic energy in SpMT processors in order to achieve a better performance to power ratio. Dynamic instruction criticality information can be effectively applied to control aggressive speculation in SpMT processors. In this paper, we present a model of micro-execution for SpMT processors to determine dynamic instruction criticality. We also present two novel techniques utilizing criticality information, namely delaying non-critical loads and criticality based thread-prediction for reducing useless computation and energy consumption. Our experiments show 17.71% and 11.63% reduction in dynamic energy for criticality based thread prediction and criticality based delayed load scheme respectively while the corresponding improvements in dynamic energy delay products are 13.93% and 5.54%.
Weitere Kapitel dieses Buchs durch Wischen aufrufen
- Criticality Driven Energy Aware Speculation for Speculative Multithreaded Processors
- Springer Berlin Heidelberg