Proactive Peak Power Management for Many-core Architectures

While power has long been a well-studied problem, most dynamic power reduction techniques, e.g., V/f scaling, clock gating, etc., exploit slack in the execution behavior of programs to reduce average power. Peak power is often left untouched. However, peak power plays a large role in determining the characteristics and hence the cost of the power supply, thermal budgeting for the chip, as well as the reliability qualification of the processor. This paper proposes proactive peak power management policies that attempt to prevent the power of a processor from exceeding a certain threshold. The threshold is chosen to be close to the peak power of the processor, thereby minimizing the inefficiency due to the growing gap between average power and peak power of a processor, especially a multi-core processor [14]. We demonstrate that proactive peak power management can enable the placement of several more cores on a die than the power budget would allow. This can result in significant (up to 47%, 33% on average) improvements in throughput for a given power budget. We also show that proactive peak power management does not have to be centralized and heavyweight and can be applied even to many-core architectures (processors with a large number of cores). We investigate a number of efficient, decentralization techniques – e.g., mapping the proactive peak power management problem to a disjunctively constrained 0-1 knapsack problem, using machine learning and classical search/optimization approaches to reduce the decision space, and using distributed control algorithms for decentralized decision making. Several of these techniques can be used even to reduce average power through traditional dynamic and global power management.