AN EFFICIENT ADAPTIVE PREDICTIVE LOAD BALANCING METHOD FOR DISTRIBUTED SYSTEMS

When allocating processors to processes in a distributed system, load balancing is a main concern of designers. By its implementation, system performance can be enhanced by equally distributing the dynamically changing workload and consequently user expectation are improved through an additional reduction on mean response time. In this way, through process migration, a rational and equitable use of the system computational power is achieved, preventing degradation of system performance due to unbalanced work of processors. This article presents an Adaptative Predictive Load Balancing Strategy (APLBS), a variation of Predictive Load Balancing Strategy (PLBS) reported elsewhere [1]. As PLBS, APLBS is a sender initiated, prediction-based strategy for load balancing. The predictive approach is based on estimates given by a weighted exponential average [12] of the load condition of each node in the system. The new approach tries to minimise traffic en the network selecting the most suitable subset of candidates to request migration and the novel aspect is that the size of this subset is adaptative with respect to the system workload. APLBS was contrasted against Random (R), PLBS and Flexible Load Sharing (FLS) [7] strategies on diverse scenarios where the load can be characterised as static or dynamic. A comparative analysis of mean response time, acceptance hit ratio and number of migration failures under each strategy is reported