Analysis of Scalable Algorithms for Dynamic Load Balancing and Mapping with Application to Photo-realistic Rendering

Abstract

This thesis presents and analyzes scalable algorithms for dynamic load balancing and mapping in distributed computer systems. The algorithms are distributed and concurrent, have no central thread of control, and require no centralized communication. They are derived using spectral properties of graphs: graphs of physical network links among computers in the load balancing problem, and graphs of logical communication channels among processes in the mapping problem. A distinguishing characteristic of these algorithms is that they are scalable: the expected cost of execution does not increase with problem scale. This is proven in a scalability theorem which shows that, for several simple disturbance models, the rate of convergence to a solution is independent of scale. This property is extended through simulated examples and informal argument to general and random disturbances. A worst case disturbance is presented and shown to occur with vanishing probability as the problem scale increases. To verify these conclusions the load balancing algorithm is deployed in support of a photo-realistic rendering application on a parallel computer system based on Monte Carlo path tracing. The performance and scaling of this application, and of the dynamic load balancing algorithm, are measured on different numbers of computers. The results are consistent with the predictions of scalability, and the cost of load balancing is seen to be non-increasing for increasing numbers of computers. The quality of load balancing is evaluated and compared with the quality of solutions produced by competing approaches for up to 1,024 computers. This comparison shows that the algorithm presented here is as good as or better than the most popular competing approaches for this application. The thesis then presents the dynamic mapping algorithm, with simulations of a model problem, and suggests that the pair of algorithms presented here may be an ideal complement to more expensive algorithms such as the well-known recursive spectral bisection.

Files

Additional details