Klaszteres terhelés-elosztó szolgáltatáspárhuzamos alkalmazásokhoz

Tóth Márton <tml@inf.u-szeged.hu>

SZTE Számítástudományi Tanszék

The large cost of process migration (we focus on process-based load balancing), the relatively small number of hosts and (nearly) uniform network connections in a cluster suggest implementing iterative algorithms with medium time complexity. The algorithms found in the literature of load balancing work on abstract processes represented as pairs of computation and communication cost values, so we need to examine the question of how to estimate these values from the actual, measured CPU and network usage values. There can be serious limitations on the validity of the above estimated values when the processes are rearranged. Because of that, we need to modify the original load-balancing algorithms, and some other modifications are made to increase their speed. The implemented algorithms are built into the P-GRADE parallel application development system, so empirical analysis of the algorithms becomes possible. We use large parallel applications for the purpose of comparing load balancing algorithms.