Energy Consumption Estimation in Parallel Applications: an Analysis in Real and Theoretical Models

Abstract

This paper presents a detailed energy consumption analysis, considering the energy consumption related to CPU, cache memory and main memory of parallel applications on HPC systems. Furthermore, this paper also presents the correlation between energy consumption, Speedup, and execution time. Experiments are conducted with the NAS parallel benchmarks using three different measurement tools: 1) Intel PCM, 2) Perf Linux, and 3) HP CACTI. The results show a comparison between two approaches to obtain energy consumption results. One using PCM and other using Perf and CACTI. The DRAM results show an average variation between these approaches of 47% for sequential applications, and 19% for parallel applications. The system results show that the lowest energy consumption occurs only when all physical cores are used, showing that the hyper-threading system did not bring benefits in energy consumption to the system. Moreover, the cache memories results show that the cache miss rate (regardless of the level) increases with the number of threads. However, a parallel application has lower cache memory energy consumption when compared to its sequential version.