A Multithreaded System for Adaptive Load Balancing in Distributed and Parallel Architectures

In this paper we present a thread-based adaptive load balancing system for distributed and parallel architec-tures. The system is composed of three agents : an information agent, a transfer agent and a location agent. The information agent uses a combination of relative and periodic information exchange to maintain a current load state in the machine. The period is adap-tive. In order to avoid the collapse of the machine this period is lowered. Its initial and lower values are determined in accordance with the target architecture of the execution. The location agent integrates a technique which permits to avoid the ood of nodes. To minimize the overhead induced by the load balancing system we used a thread-based implementation. An application of the system to an IDA program applied to the 15-puzzle problemm4] is presented. Some preliminary results obtained on a farm of DEC/ALPHA processors under PM 2 6] are exposed.