Slowdown-Guided Genetic Algorithm for Job Scheduling in Federated Environments

Large-scale federated environments have emerged to meet the requirements of increasingly demanding scientific applications. However, the seemingly unlimited availability of computing resources and heterogeneity turns the scheduling into an NP-hard problem. Unlike exhaustive algorithms and deterministic heuristics, evolutionary algorithms have been shown appropriate for large-scheduling problems, obtaining near optimal solutions in a reasonable time. In the present work, we propose a Genetic Algorithm (GA) for scheduling job-packages of parallel task in resource federated environments. The main goal of the proposal is to determine the job schedule and package allocation to improve the application performance and system throughput. To address such a complex infrastructure, the GA is provided with knowledge based on slowdown predictions for the application runtime, obtained by considering heterogeneity and bandwidth issues. The proposed GA algorithm was tuned and evaluated using real workload traces and the results compared with a range of well-known heuristics in the literature.