LearnPower: Exploiting Program Similarity for Learning based Power Management in Many-Core Processors

LearnPower tackles the power management challenge in many-core processors, which has become the limiting factor for continuing microprocessor roadmap.

It proposes a novel learn & infer power management paradigm that exploits program similarity to learn and infer most efficient power provisioning.

Off-line learning is devised to find the explicit function between power and application performance, and then on-line inference is used to infer the optimal power operating point which maximizes performance-power efficiency.LearnPower is in sharp contrast to incumbent solutions failing into either the profile & supply or the measure & tune paradigm.

The profile & supply paradigm is to "provide power on demand" by profiling hardware activities (e.g. core idle time, buffer backlog, packet intensity and criticality) as a proxy of need and then supplying voltage/frequency levels according to the need.

The measure & tune paradigm is to "provide power for performance target" by measuring performance indicative metrics (e.g. memory access latency, packet delay, network throughput) and then tuning voltage/frequency levels according to intended performance target.Power management in many-core processors is a must-to-do but has no more "low-hanging fruits".

LearnPower aims for a paradigm shift to pursue extreme power provisioning efficiency, which cannot be otherwise achieved by existing approaches. It is proposed for the PI (20%) and one PhD student (80%) for four years.