A prediction-based adaptive location policy for distributed load balancing

Abstract

The location policy in distributed load balancing is to locate the destination nodes to or from which tasks will be transferred. An efficient location policy is required to achieve high performance on distributed load balancing. In this paper, we propose a new distributed adaptive location policy based on predictable stare knowledge. The predictable state knowledge in each node is composed of the system state information collected at run time and the predefined static information that is a priority order of each node for transferring tasks. The proposed scheme systematically maintains the predictable state knowledge in each node by using an efficient data structure and a rule for collecting state information with low overheads. When the state of a node becomes heavily-loaded, the proposed scheme predicts both heavily-loaded nodes and lightly-loaded nodes by exploiting predictable state knowledge and then finds a good lightly-loaded node that minimizes useless polling and maximizes even load distribution. An analytic model is developed to compare the presented scheme with other well known schemes. The validity of the model is checked with an event-driven simulation, and it is shown that the presented scheme exhibits a significant performance improvement over other schemes, especially at high system loads. Also, the presented scheme is shown to significantly improve polling hit ratio and to avoid system instability.