PGCapping
Kai Ma, Xiaorui Wang · 2012
Optimizing the performance of a chip multiprocessor (CMP) within a power cap has recently received a lot of attention. However, most existing solutions rely solely on DVFS, which is anticipated to have only limited actuation ranges in the future. Power gating shuts down idling cores in a CMP, such that more power can be shifted to the cores that run applications for better CMP performance. However, current preliminary studies on integrating the two knobs focus on deciding the power gating and DVFS levels in a tightly coupled fashion, with much less attention given to the direction of decoupled designs. By decoupling the two knobs that may interfere with each other, individual knob management algorithms can be less complex and more efficient to take advantage of the characteristics of different knobs. This paper proposes PGCapping, a decoupled design to integrate power gating with DVFS for CMP power capping. To fully utilize the power headroom that is reserved through power gating, PGCapping enables per-core overclocking on turned-on cores that run sequential applications. However, per-core overclocking may make some cores age much faster than others and thus become the reliability bottleneck in the whole system. Therefore, PGCapping also uses power gating to balance the core lifetimes. Our empirical results on a hardware testbed show that the proposed scheme achieves up to 42.0% better average application performance than five state-of-the-art power capping baselines for realistic multi-core applications, i.e., a mixed group of PARSEC and SPEC CPU2006 benchmarks. Furthermore, our extensive simulation results with real-world traces demonstrate that a lightweight lifetime balancing algorithm (based on power gating) can increase the CMP lifetime by 9.2% on average.