Workload Generation for Trace Based Grid Simulations

Testing Grid models and their resource management components is a critical part of efficient Grid development. There are two ways to experimentally assess the performance of Grid models: benchmarking and using synthetic grid workloads. Workload generation for Grids has become more problematic than for independent supercomputers. It requires standardization of the activity of thousand simultaneous users that use a broad amount of different resources. In this paper, we address the problem of synthetic Grid workload generation, and propose strategies based on real production traces. Carefully reconstructed traces from real supercomputers can provide a very realistic job stream for simulation-based performance evaluation of Grid job scheduling algorithms. We consider two types of jobs that can be run in Grids, and therefore included into grid workloads: unitary jobs such as sequential and parallel (e.g., MPI) applications, and composite jobs (workflow applications) that represent executable patterns of process interactions, and describe the execution of a complex application built from individual application components.