FEPAC: A Framework for Evaluating Parallel Algorithms on Cluster Architectures

Abstract

For many years, computer scientists have explored the computing power of so-called computing clusters to address performance requirements of computationally intensive tasks. Historically, computing clusters have been optimized with run-time performance in mind, but increasingly energy consumption has emerged as a second dimension that needs to be considered when optimizing cluster configurations. However, there is a lack of generally available tool support to experiment with cluster and algorithm configurations in order to identify “sweet-spots” with regards to both, run-time performance and energy consumption, respectively. In this work, we are introducing FEPAC, a framework for the automated evaluation of parallel algorithms on different cluster architectures and different deployments of software processes to hardware nodes, allowing users to explore the impact of different configurations on run-time properties of their computations. As proof of concept, the utility of the framework is demonstrated on a custom-built Raspberry Pi 3B+ cluster using different types of parallel algorithms as benchmarks. The experiments evaluate matrix multiplication, kmeans, and OpenCV on varying sizes of cluster, and showed that although a larger cluster improves performance, there is often a trade-off between energy and computation time.