In this paper we explore the tradeoff between the leakage power and fan power to dynamically migrate tasks to minimize the overall power consumption in a homogeneous many-core processor. Our analysis shows that the overall power can be minimized if a task allocation for minimum peak temperature is adopted together with an intelligent fan speed adjustment that finds the optimal tradeoff between fan power and leakage power. We propose a method to compute the lower bound on the minimum peak temperature among all possible allocations of given a task set. We further propose two global heuristic task mapping algorithms and a multi-agent distributed task migration framework that minimizes the peak temperature during runtime. The proposed framework achieves large fan power saving as well as overall power reduction. Experimental results show that, given a tight temperature constraint, our distributed task migration policy can save up to 38.5% fan power and 28.9% overall system power compared to the best random mapping policy. Our data also show that the overall system power is insensitive to the task allocation when the temperature constraint is loose.
- Low power
- Multi-agent distributed framework
- Task allocation
- Thermal aware
ASJC Scopus subject areas
- Electrical and Electronic Engineering