Analysis of coordinated load sharing for large distributed systems

The performance of a distributed system (DS) is limited by how good the load sharing scheme used in the system is. This fact is especially true for a large-scale DS where the communication bandwidth between machines may be limited. Load sharing is the process of sharing computational resources by distributing the load of a distributed computing system among the various computers in the system. This paper proposes and analyzes new type of algorithms for sharing system loads based on a coordinated dynamic semi-centralized scheme. In this scheme, load state information at each node is centrally collected by a coordinator to optimize the transfer policy, a policy for transferring loads from one computer to another. This information is periodically broadcasted to nodes for decentralizing the decision making for best load allocation policy. The paper discusses the implementation of the scheme and the optimization of its parameters for small systems as well as for large DS where the system is partitioned into multiple partitions according to the available communication bandwidth. Simulations results have shown that this coordinated scheme outperforms widely-used probing-based schemes in most system loads.