Large-Scale Parallel Computing on Grids

This paper argues that computational grids can be used for far more types of applications than just trivially parallel ones. Algorithmic optimizations like latency-hiding and exploiting locality can be used effectively to obtain high performance on grids, despite the relatively slow wide-area networks that connect the grid resources. Moreover, the bandwidth of wide-area networks increases rapidly, allowing even some applications that are extremely communication intensive to run on a grid, provided the underlying algorithms are latency-tolerant. We illustrate large-scale parallel computing on grids with three example applications that search large state spaces: transposition-driven search, retrograde analysis, and model checking. We present several performance results on a state-of-the-art computer science grid (DAS-3) with a dedicated optical network.