Distributed Systems Using Mobile Agents

In this paper, we present a new load balancing algorithm for parallel virtual architectures based on mobile agents. The proposed model is based on a reference diagnostic test performed by the host agent at the deployment time of the parallel application. This diagnostic test allows evaluating the computation performance and the communication latency of each node of the distributed system. The parallel tasks to execute are assigned to the distributed virtual processing units (VPUs) according to the performance offered by each node. The VPUs, represented by mobile agents, are associated in a parallel virtual layer, according to the parallel architecture topology required by the application. This model can also use metadata, of the parallel tasks to be distributed, to estimate more accurately the required times to perform the current parallel task. This feature allows predicting the performance offered by the distributed system according to the selected configuration of the selected computational model. When the system detects unbalanced loads, during the execution of any iterative computation, the system determines the required migrations of the VPUs from overloaded to under loaded nodes in order to keep the distributed system in a perfect balanced stat. To describe the proposed approach, the established mathematical model and a SPMD parallel application example are presented to prove the effectiveness of the proposed model.