A NEW FRAMEWORK FOR DYNAMIC LOAD BALANCING IN DISTRIBUTED COMPUTER SYSTEMS

Abstract

A number of/rends in computational science and engineering have increased the need for effective distributed computer system. Effective utilization of distributed computing ,system requires effrcie7rt policies for resource allocation. Sharing r)f computing power is usually done in the firm (?f load (task) sharing or load balancing. Thus load balancing is useful in attempting to keep all processors busy if any job is waiting. In a distributed computer system, tasks arrive at the processors in a random . fashion with unpredictable load estimates, so it is likely that the load balancing is conducted by transferring some jobs from processors that are heavily loaded to those that are lightly loaded or idle for processing. Thus balancing the workload over a distributed computer system is important to improve its overall performance. This dissertation presents a dynamic load balancing framework for multiclass jobs in distributed computer systems. Strategies for dynamic load balancing are discussed f )r two multiprocessor networks, viz., Hypercube and Folded mesh structure. As part ?f the work, first an adaptive heat diffusion scheme is presented. Next, the task selection mechanism given therein, has been modified to improve communication locality. Simulation study is carried out for performance evaluation of these strategies on n-dimensional hypercrrhe structure. Results show that the algorithm can obtain a nearly optim7.cnr solution for a given communication overhead and response time. The simulation program for this thesis is written in 'C' language and development environment is IRIX on TA TA ELXV POWER SERIES 3200 system.