Multi-agent Model of Resource Allocation in Distributed Computing Environment for Job Flow Processing

Currently, a resource allocation issue in a high-performance computing environment is still a challenge. Its solution enables us to improve large spectrum criteria. Among them are the time, cost, reliability and other parameters of problem-solving by end-users. High computational complexity related to optimal solving this issue through the hard search among all scenarios of the resource allocation. At the same time, we can apply knowledge-based heuristic approaches to obtain satisfactory sub-optimal solutions. Applying a multi-agent technology is one such approach. In the paper, we propose a multi-agent model of resource allocation in a high-performance computing environment in the job flow processing. We suppose that such job flows are generated by distributed applied software package. A high-performance computing environment is based on the resources of a public access supercomputer center. In such conditions, we compare traditional meta-schedulers (GridWay and Condor DAGMan) and our multi-agent system in terms of the effectiveness of job flow management. We provide the experimental results that obviously show advantages of the proposed multi-agent management in speedup and efficiency of computations, processor load, and its balancing.