Estimating the Energy Consumption of Applications Running on Heterogeneous Machines

Parallel and distributed platforms available today are becoming more and more \textit{heterogeneous}. Cluster and Supercomputer systems are built exploiting this heterogeneity using CPUs, GPUs, FPGAs and other accelerators. The goal in a heterogeneous environment is to utilize different resources to enable the accomplishment of a given task in the shortest possible time and lately energy consumption. This post-doctoral project aims to address the environmental impact of scientific workflow executions on heterogeneous supercomputers by developing advanced scheduling techniques that optimize for both performance and sustainability. Scientific Workflows (SWs) are computational applications built with tasks from different operations that form a directed acyclic graph (DAG). The core objective of this research is to reduce carbon footprint associated with supercomputing tasks by enhancing scheduling algorithms through multi-objective reinforcement learning (MORL). By incorporating MORL, the scheduling process will simultaneously consider multiple objectives, such as minimizing energy consumption, reducing carbon emissions, and maintaining high computational efficiency. This research aims to reduce greenhouse gas emissions associated with the execution of scientific workflows. By utilizing regression techniques, we will predict the performance and energy consumption of tasks in various scientific workflows. This information will then be used to develop intelligent, energy-aware schedulers using novel MORL-based scheduling algorithms.\\ \textbf{General Objective:} The primary objective of this project is to reduce greenhouse gas emissions associated with the execution of scientific workflows in heterogeneous supercomputers. This will be achieved by proposing improvements to scheduling techniques with with multi-objective reinforcement learning in scientific workflow management systems, leveraging predictions of performance and energy consumption across heterogeneous resources. \\ \textbf{Key Words:} Distributed Systems, Heterogeneity, Schedulers, Performance Analysis, Machine Learning, Scientific Workflows, multi-objective Reinforcement Learning. \\