Dynamic Core Allocation for Energy-Efficient Thread-Level Speculation

Thread-level speculation becomes promising for allowing multiple dependent threads to run on a multi-core processor simultaneously. It is often assumed that the parallel performance of a program is increased linearly as the number of processor cores increases. Due to ambiguous control and data dependences, the performance is either limited by frequent dependence violation, or achieved at the cost of energy consumption. Thus we present an approach to dynamically determine an appropriate number of cores for each region and make an energy-efficient speculation. It relies on the runtime performance profiles of speculative threads to reveal the effect of thread interference on the speculation. This effect is classified into positive and negative two classes due to their impact on performance. Both of them are then estimated by the order of speculation. Based on such estimation, we dynamically leverage the energy-efficiency of speculative threads at each invocation and choose the right core count for them to parallelize. We have evaluated our approach using SPEC CPU2000 benchmarks. It not only achieves comparable or better performance, but also largely reduces the energy consumption.