Costs and benefits of load sharing in the computational grid

We present an analysis of the costs and benefits of load sharing of parallel jobs in the computational grid. We begin with a workload generation model that captures the essential properties of parallel jobs and use it as input to a grid simulation model. Our experiments are performed for both homogeneous and heterogeneous grids. We measured average job slowdown with respect to both local and remote jobs and we show that, with some reasonable assumptions concerning the migration policy, load sharing proves to be beneficial when the grid is homogeneous, and that load sharing can adversely affect job slowdown for lightly-loaded machines in a heterogeneous grid. With respect to the number of sites in a grid, we find that the benefits obtained by load sharing do not scale well. Small to modest-size grids can employ load sharing as effectively as large-scale grids. We also present and evaluate an effective scheduling heuristic for migrating a job within the grid.