MultiObjective Optimization to Workflow Grid Scheduling using Reference Point based Evolutionary Algorithm

Abstract

Listen

Translate article

ABSTRACT Grid facilitates global computing infrastructure for user to consume the services over the network. To optimize the workflow grid execution, a robust multi-objective scheduling algorithm is needed. In this paper, we considered three conflicting objectives like execution time (makespan), total cost and reliability. We propose a multi-objective scheduling algorithm, using R-NSGA-II approach based on evolutionary computing paradigm. Simulation results shows that the proposed algorithm generates multiple scheduling solutions near the Pareto optimal front with small computation overhead. General Terms Distributed systems Keywords Workflow Grid Scheduling, Multi-objective Optimization, MOEA, Pareto dominance. 1. INTRODUCTION Grid computing technologies primarily emerged to become the next generation of high performance computing by placing numerous heterogeneous resources of many organizations. Scheduling in Grid computing is the hot topic of research and challenging due to heterogeneity and dynamism of resources in grid. In this paper, we consider Directed Acyclic Graph (DAG), an application model for describing workflow. Scheduling of workflows in grid allows mapping of tasks on heterogeneous resources according to a set of procedural rules. Dynamism of resources in grid is an important issue while making scheduling decisions, in which resources can fail inevitably. Failures of resources have adverse effects on performance of workflow application. Therefore, an effective scheduling algorithm should consider the failure rate of resources in order to make maximum reliability of the schedule. Scheduling is the NP-hard problem; so many heuristic approaches have been applied in the grid workflow. One of the primary motives of any grid system is to meet user requirements in an intuitive way by considering multiple objectives or criterion. Many different criterion can be considered in scheduling of complex workflow computational tasks, usually include execution time of the task, cost of the task to run on a resource, utilization of resources, reliability, turnaround time and many others. In the recent years, many heuristics have been applied in scheduling of grid in the consideration of single criteria and pairs of certain criterion to generate single solution to the users but failed to fully satisfy users. For maximum satisfaction of user, it is necessary to produce multiple solutions with respect to minimization or maximization of objectives and selection of a solution from these solutions is further left to the user. Thereby, an optimization of conflicting multiple objectives is required to generate multiple tradeoff solutions. The Multi-Objective Evolutionary Algorithms (MOEAs) are the effective way to solve multi-objective optimization problem like scheduling in grid. An MOEA approach produces Pareto optimal set of solutions, which is the set consisted of all non-dominated solutions. A solution is called non-dominated solution if it is at least best in one objective with respect to others. This paper focuses on three major conflicting objectives namely execution time (makespan), cost and reliability of the schedule in order to generate schedules under deadline and budget constraints specified by the user. According to scheduling problem, both execution time and cost are minimization objectives while reliability is the maximization objective. But we consider reliability using reliability index as minimization objective. Since, we are interested in the preference set of solutions near the user specified region of interest. Towards this goal we considered reference point based non dominated sort genetic algorithm (R-NSGA-II). Rest of the paper is organized as follows. Section 2 specifies some of the related work. In section 3, we introduced the Grid Workflow Scheduling problem definition. Section 4, describes the technique of multi objective optimization and different multi objective evolutionary algorithms used. Section 5 discusses the simulation analysis of MOEA approaches used. Finally section 6 gives the conclusion.