A lightweight Java taskspaces framework for scientific computing on computational grids

TL;DR: The scalable Taskspaces framework runs fully transparently on heterogeneous grids while maintaining a low complexity in terms of installation, maintenance, application programming and grid operation and offers a promising roadway to push scientific grid computing with intertask communication beyond the experimental research setting.

Abstract: A prototype Taskspaces framework for grid computing of scientific computing problems that require intertask communication is presented. The Taskspaces framework is characterized by three major design choices: decentralization provided by an underlying tuple space concept, enhanced direct communication between tasks by means of a communication tuple space distributed over the worker hosts, and object orientation and platform independence realized by implementation in Java. Grid administration tasks, for example resetting worker nodes, are performed by mobile agent objects. We report on large-scale grid computing experiments for iterative linear algebra applications showing that our prototype framework scales well for scientific computing problems that require neighbor-neighbor intertask communication. It is shown in a computational fluid dynamics simulation using a Lattice Boltzmann method that the Taskspaces framework can be used naturally in interactive collaboration mode. The scalable Taskspaces framework runs fully transparently on heterogeneous grids while maintaining a low complexity in terms of installation, maintenance, application programming and grid operation. It thus offers a promising roadway to push scientific grid computing with intertask communication beyond the experimental research setting.