Machine Learning Based DVFS for Energy Efficient Execution of Multithreaded Workloads
MetadataShow full metadata
Concerns over high power consumption of large computations and data centers have been growing in recent years. Many software and hardware strategies for reducing power have been proposed as remedies. Dynamic voltage and frequency scaling (DVFS) is one technique that can be effective if given expert knowledge. However, DVFS effectiveness is sensitive to workload characteristics and architectural parameters. Lack of knowledge can hurt DVFS strategies and render it ineffective. This thesis presents a supervised machine learning (ML) strategy for automatically making smart DVFS decisions to improve energy efficiency of multi threaded and multiprogram workloads. The technique uses hardware performance counters to construct feature vectors that capture program behavior and thread interaction in a meaningful way. The resulting models have high accuracy in picking optimal frequencies. Experimental results on contemporary benchmark suite show that application of a ML technique is able to reduce energy consumption by as much as 24% on memory-intensive workloads.