On Grid Performance Evaluation Using Synthetic Workloads

摘要

Grid computing is becoming a common platform for solving large scale computing tasks. However, a number of major technical issues, including the lack of adequate performance evaluation approaches, hinder the grid computing's further development. The requirements herefore are manifold; adequate approaches must combine appropriate performance metrics, realistic workload models, and flexible tools for workload generation, submission, and analysis. In this paper we present an approach to tackle this complex problem. First, we introduce a set of grid performance objectives based on traditional and grid-specific performance metrics. Second, we synthesize the requirements for realistic grid workload modeling, e.g. co-allocation, data and network management, and failure modeling. Third, we show how GrenchMark, an existing framework for generating, running, and analyzing grid workloads, can be extended to implement the proposed modeling techniques. Our approach aims to be an initial and necessary step towards a common performance evaluation framework for grid environments.