Using machine learning in trace-driven energy-aware simulations of high-throughput computing systems.

Abstract

When performing a trace-driven simulation of a High Throughput Computing system we are limited to the knowledge which should be available to the system at the current point within the simulation. However, the trace-log contains information we would not be privy to during the simulation. Through the use of Machine Learning we can extract the latent patterns within the trace-log allowing us to accurately predict characteristics of tasks based only on the information we would know. These characteristics will allow us to make better decisions within simulations allowing us to derive better policies for saving energy. We demonstrate that we can accurately predict (up-to 99% accuracy), using oversampling and deep learning, those tasks which will complete while at the same time provide accurate predictions for the task execution time and memory footprint using Random Forest Regression.