Dynamic allocation of communicating tasks in computational grids

This paper studies task allocation in computational grids operating in a dynamic and uncertain environment. Computational grids consist of loosely coupled heterogeneous resources or agents with finite buffer capacities. These grids are primarily used to process large-scale applications consisting of several interdependent tasks. The task allocation problem is modeled as an infinite horizon Markov decision process, with the resource service times and the task arrivals following general probability distributions. We explicitly consider the communication cost between agents incurred in coordinating the execution of interdependent tasks. We show that a stationary optimal policy exists for this task allocation problem. Furthermore, we develop an action elimination procedure for reducing the complexity of computational methods in finding the optimal policy. We also present a real-time heuristic policy based on certain structural properties of the problem. Finally, computational results are presented that compare the performance of the heuristic policy with respect to other approaches for allocating tasks in the grid. Results from this paper are also applicable to the task allocation problem in manufacturing and service areas such as distributed design, project management and supplier coalitions.