Towards a predictive load balancing method based on multiples resources

Abstract

Processors load unbalance in distributed systems is one of the main problems, because ít involves system performance degradation. Load balance algorithms try to improve the system global performance through migration of processes, but they present also an additional problem, known as instability: lt happens when processes spend an excessive amount of time migrating among different system nodes. In arder to diminish this cost without affecting the mean system response time, load balancíng algoríthms based on dífferent strategíes have been proposed. Multiple Resources Predictíve Load Balance Strategy (MRPLBS), ís a new predíctive, dynamic and nonpreemptive strategy for balancing multiple resources. The predictive approach is based on estimations computed as weighed exponential averages of the load of each node in the system. This paper presents MRPLBS' system architecture and its performance and system a comparison on different scenarios against Random Load Balancing. The number of requirements, the mean response time, the number of failed migratíons and the percentage of acceptance are shown