PowerTrader: Enforcing Autonomous Power Management for Future Large-Scale Many-Core Processors

Abstract

Existing power management approaches for modern many-core processors resort to “centralized” design concept, aiming to optimize chip performance under fixed power budget. Unfortunately, the centralized power management approach, which usually relies on a dedicated on-chip power manager, faces various limitations such as poor scalability and high implementation overhead, and hence cannot be deployed in future large-scale manycores. This article proposes PowerTrader, an autonomous power management scheme. PowerTrader endows each core with self autonomy to issue the power control at any time to harvest the desirable power quota through negotiating with vicinity cores. It does not incur the overheads introduced by power allocation and statistics collection that are inevitable in centralized approaches, meanwhile chip power consumption could be well kept beneath the preset power budget. This article also elaborates on the key design tradeoff in autonomous power management (i.e., Mean-Time-to-Stable versus application power efficiency), and provides thorough design space exploration to justify the efficacy of the proposed approach. Experimental results show that PowerTrader achieves substantial improvements in both performance and power, and exhibits superior scalability compared with the state-of-the-arts.