How to Design a Job Scheduling Algorithm

Abstract

We discuss design aspects of job scheduling algorithms in parallel processing systems. Starting from the observation that in this area the impact of most research publications on real systems is negligible, we first identify three main categories with strong relevance to practical implementations: constraints, objectives, and evaluation. For each category, we describe common aspects of real systems that are presently important and then introduce some general rules that should be followed when presenting a new job scheduling algorithm for parallel processing systems. To apply these rules, we use as an example a new method for the Infrastructure-as-a-Service (IaaS) model of Cloud Computing that extends the spot instance approach of Amazon. In particular we address technical, organizational, and usage constraints based on existing concepts and research results for this example. Then we briefly describe scheduling objectives from the viewpoint of a data center offering IaaS. After presenting our algorithmic concept, we provide an evaluation with theoretical means to demonstrate that this approach can also yield meaningful results in practice.