Scheduling in Heterogeneous Distributed Systems

Abstract

Parallel and distributed systems play an important part in the improvement of high-performance computing. In analyzing the performance of such Heterogeneous Distributed Systems (HDS), scheduling a set of tasks to the available set of resources for execution is highly important. Task-scheduling being an NP-complete problem, use of meta-heuristics is more appropriate in obtaining optimal solutions. The problem of scheduling an application comprising a set of independent tasks on a fully connected HDS is considered here. The scheduling algorithms for this problem focus on minimizing two objectives, the make-span and the flow-time. These objectives conflict with one another, which requires multi-objective problem formulation. Multi-objective Genetic Algorithms (MOGAs) and Multi-objective Particle Swarm Optimization (MOPSO) algorithms are applied to the problem. Weighted Sum Genetic Algorithm (WSGA) based on weighted sum approach, and Nondominated Sorting Genetic Algorithm (NSGA2), a modified version of NSGA2 with controlled elitism (NSGA2-CE) and a hybrid version of NSGA2 combined with Pareto hill climbing (PHC), the NSGA2-PHC are applied to the problem. A weighted sum particle swarm optimization (WSPSO), a Nondominated Sorting Particle Swarm Optimization (NSPSO), an Adaptive Nondominated Sorting Particle Swarm Optimization (ANSPSO), and a hybrid version of ANSPSO with PHC, the NSPSO-PHC are the MOPSO methods applied to task scheduling. The different MOGA and MOPSO methods are compared among themselves to determine the algorithms that generate the efficient schedule. The best version of MOGA is compared with the best MOPSO method to find that MOGA produces better schedules for benchmark test instances simulated.