On the Competitive Analysis of Randomized Static Load Balancing

Static load balancing is attractive due to its simplicity and low communication costs. We analyze under which circumstances a randomized static load balancer can achieve good balance if the subproblem sizes are unknown and choosen by an adversary. It turns out that this worst case scenario is quite close to a more specialized model for applications related to parallel backtrack search. In both cases, a large number of subproblems has to be generated in order to make good load balance possible. Nevertheless, a carefully implemented randomized static load balancer can sometimes compete with dynamic load balancing on parallel machines with slow communication. The ideas and results derived here can also be used to analyze and improve existing load balancing algorithms.