Evaluation of a Practical Markov model-based Methodology for Energy Efficiency in Multicore Systems

Greedy heuristic methodologies perform the dynamic voltage and frequency scaling (DVFS) on multicore systems based on the localized system performance data. The characteristics of applications such as their floating-point operations and memory loads can vary largely during the runtime. As such, the heuristic-based prediction of Voltage and Frequency (V/F) is suboptimal and leads to performance degradation in applications with widely varying computational demands. This work develops a new approach, Markov-based DVFS (MDVFS), which formulates a Hidden Markov Model (HMM) for predicting core utilizations (states) and V/F levels based on the observable execution times of cores across application time intervals. MDVFS uses the well-known K-means algorithm for defining states and observation levels of HMM. For solving the HMM problem, MDVFS uses the Viterbi algorithm to find a sequence of V/F levels that most likely generate the execution time sequence. This paper uses Energy-Delay
²
Product (ED
²
P) for measuring system energy efficiency and compares the performance of MDVFS to 2 predictive algorithms, history and feedback controller, on multiple applications. Experimental results show that for 5 out of 6 applications MDVFS, on the average, obtains up to 34% faster execution runtime and improves ED
²
P by 24% compared to the history and the feedback controller.