Design and Evaluation of a Parallel Data Redistribution Component for TGrid

Abstract

Abstract. Data redistribution of parallel data representations has become an important factor of grid frameworks for scientific computing. Providing the developers with generalized interfaces for flexible parallel data redistribution is a major goal of this research. In this article we present the architecture and the implementation of the redistribution module of TGrid. TGrid is a grid-enabled runtime system for applications consisting of cooperating multiprocessor tasks (M-tasks). The data redistribution module enables TGrid components to transfer data structures to other components which may be located on the same local subnet or may be executed remotely. We show how the parallel data redistribution is designed to be flexible, extendible, scalable, and particularly easy-to-use. The article includes a detailed experimental analysis of the redistribution module by providing a comparison of throughputs which were measured for a large range of processors and for different interconnection networks.

Dynamic scheduling of multi-processor tasks on clusters of clusters

“…Dependencies between tasks are modeled as edges. Data dependencies are automatically resolved by TGrid [7], i.e. the redistribution component queries the sending (preceeding) and receiving (succeeding) task for the format of input/output data (size, type, etc.)…”