Leakage Power Aware Task Assignment Algorithms for Multicore Platforms

Abstract

Increased power density and high temperatures are looming issues in many-core processors. Technology scaling trends, cooling limitations, and stringent application requirements make these issues rather difficult to handle. Consequently, it is imperative to design solutions that are effective and also scale well with increasing core counts. In this work, we present an application mapping framework LeakOpt, which aims to minimize the total power consumption of manycore processors. We first demonstrate the implications of lateral heat conduction on leakage power consumption and show that heat spread aware task assignment can significantly impact the total power consumption. We formulate the mapping problem as an optimization problem and design a family of algorithms to solve it heuristically. We present simulation results that shows reduction upto 27.12% in leakage power consumption relative to worst case task mapping for a variety of workloads. Heuristic based mapping schemes perform 2600x faster (for 225 cores) while still within 2.5% of best case results. We further evaluate the same algorithms on a real hardware (TILE-Gx36TM) and show that these techniques can reduce leakage by upto 18.22% on average. Results on hardware are consistent with the simulation results as far as the relative effectiveness of various heuristics is concerned.