Supporting High-Level Grid Parallel Programming: the Design and Implementation of Grid-GUM2

Computational Grids are much more complex than classical high-performance architectures: they have high, and dynamically varying communication latencies, and comprise heterogeneous collections of processing elements. We argue that such a complex and dynamic architecture is extremely challenging for the programmer to explicitly manage, and advocate a high-level parallel programming language, like GpH, where the programmer controls only a few key aspects of the parallel coordination. Such a high-level language requires a sophisticated implementation to manage parallel execution. This paper presents the design and implementation of Grid-GUM2, a Grid-specific runtime environment for GpH. The core of the design are monitoring mechanisms that collect static and partial dynamic information, and new load management mechanisms that use the information to better schedule tasks on computational Grids. These mechanisms are built on top of a virtual shared memory (VSM), graph-reduction-based computation engine. A systematic evaluation of the implementation, reported elsewhere, shows acceptable scaleup and good speedups on a range of computational Grids.