Dynamic load balancing in distributed heterogeneous computer systems

This thesis presents a study of load-balancing in a distributed system consisting of a number of heterogeneous hosts connected by a local area network. We motivate resource sharing by investigating imbalance and find that the probability of uneven loading is high for large systems under moderate load. In order to be able to evaluate alternate load sharing schemes, an approximate queueing technique, the decoupled iterative technique, is proposed. This methodology determines the mean response time performance of large networks with heterogeneous interactive queues. To illustrate the effectiveness and efficiency of the technique, we include simple solutions for several complicated examples. A GPSS/PC simulation is used to validate our approximate technique. Another validation method used is comparison with known solutions to special cases. We propose a new dynamic load-balancing scheme, called the enhanced receiver-initiated dynamic algorithm (ERIDA). The impact of the communication overhead is discussed. We also propose another effective scheme for load balancing, a two-mode dynamic algorithm (TMDA). The performance profile of this algorithm shows that this algorithm has better delay performance in the medium-to-high load range than either the sender-initiated scheme or the receiver initiated scheme. The iterative modelling technique and extensive simulation are applied to the two new load-balancing schemes to evaluate the mean response time performance. (Copies available exclusively from Micrographics Department, Doheny Library, USC, Los Angeles, CA 90089-0182.)