Lehrstuhl für Effiziente Algorithmen des Instituts für Informatik der Technischen Universität München Randomised Dynamic Load Balancing

Load balancing is about distributing work load (jobs, tasks, processes, etc.) among a set of processing facilities (processors, workstations, servers, etc.), such that, usually, this load is more or less evenly distributed. In this work we introduce and investigate the performance of three randomised load balancing algorithms for dynamic settings, that is, we are interested mostly in the long term behaviour of the algorithms, and here especially in deriving an upper bound on the maximum load of any server at any point of time. We assume two fundamentally different load generation schemes. First, we analyse the algorithms given a stochastical scheme, i.e., the generation and consumption of load obeys some probabilistic distribution. Second, we introduce an adversarial scheme, where we assume that the generation and consumption of load is controlled by an adversary, who deliberately tries to produce a load distribution as uneven as possible. Two of our algorithms are for the stochastical generation scheme, and the third one is for the adversarial scheme. After rigorously analysing the performance of the algorithms, we present some simulation results that indicate that under certain conditions our algorithms very are well-behaved in practice, too. Since the algorithms are somewhat theoretical in nature, we also briefly discuss how to modify them and tune certain parameters in order to make them even more applicable for practical problems.