Load balancing in peer-to-peer systems using a diffusive approach

摘要

We developed a diffusive load balancing technique for P2P systems. This technique uses the overlay network of a P2P system and results in the nodes of the network having similar available capacities; therefore the services hosted on these nodes are expected to have similar mean response times. In this paper, the technique is presented, including the policies, stages of operation, and decision algorithms. The convergence of the available capacities to the global average is demonstrated. The convergence speed depends on the decision algorithm, the neighborhood structure of the underlying overlay network, and the workload distribution. When used in a system with churn, the technique keeps the standard deviation of available capacities in the system within a bound. This bound depends on the amount of churn and the frequency of the load balancing operations, as well as on the distribution of node capacities. However, the sizes of services have little impact on this bound. The paper presents the results of analytical analysis and simulation studies.