Affiliations: School of Computer Science and Electronic Engineering, University of Essex, Colchester, UK | Department of Electrical Engineering, COMSATS Institute of Information Technology, Wah Cantt. Pakistan
Note: [] Corresponding author. Klaus D. McDonald Maier, School of Computer Science and Electronic Engineering, University of Essex, CO4 3SQ, Colchester, UK. E-mails: [email protected]; [email protected] (Muhammad Yasir Qadri); [email protected] (Nadia N. Qadri).
Abstract: Multicore architectures offer an amount of parallelism that is often underutilized, as a result these underutilized resources become a liability instead of advantage. Inefficient resource sharing on the chip can have a negative impact on the performance of an application and may result in greater energy consumption. A large body of research now focuses on reconfigurable multicore architectures in order to support algorithms to find optimal solutions for improved energy and throughput balance. An ideal system would be able to optimize such reconfigurable systems to a level that optimum resources are allocated to a particular workload and all the other underutilized resources remain inactive for greater energy savings. This paper presents a fuzzy logic based reconfiguration engine targeted to optimize a multicore architecture according to the workload requirements for optimum balance between power and performance of the system. The proposed fuzzy logic reconfiguration engine is designed around a 16-core SCMP architecture comprising of reconfigurable cache memories, power gated cores and adaptive on-chip network routers for minimizing leakage energy effects for inactive components. A coarse grained architecture was selected for being able to reconfigure faster, thus making it feasible to be used for runtime adaptation schemes. The presented architecture is analyzed over a set of OpenMP based parallel benchmarks and results show significant energy savings in all cases.
Keywords: Energy efficiency, Fuzzy logic, reconfiguration, Multicore processing
