An Improved Stochastic-based Approach for Optimizing Energy-Time Tradeoff in Multicore Systems

Abstract

Stochastic techniques are used in Dynamic Voltage and Frequency Scaling (DVFS). These techniques determine cores V/F levels based on the probability distributions of system parameters, which usually represent the computational characteristics of applications. The Hidden Markov Model (HMM) is a stochastic model, which learns the hidden states of systems based on observable events. Our previous work built an HMM for each core of a multicore system using a fixed number of states and V/F levels. The HMM parameters were used by the Viterbi algorithm to perform per-core DVFS by predicting the most likely state sequences. This paper aims at improving those state sequence predictions by examining different numbers of states. This examination impacts the size of HMM parameters that guide Viterbi for predicting the state sequences. In addition, for any number of states, the proposed model exhaustively assigns V/F levels to states to optimize Energy-Delay Product (EDP). Results show that the proposed model improves EDP, on average, by 30% compared to our previous work across multiple applications. Compared to two other predictive techniques, the proposed model obtains up to 30% more EDP improvement for applications with highly varying computations.