Implicit Information Approach for Self-Scheduling Load Sharing Policy

This paper addresses the issues of dynamic load sharing in large-scale computing systems. Typically, dynamic load sharing policies use explicit information to make load sharing decision. Our load sharing policy, however, uses consuming rate, which is implicit information, as a resource metric. This implicit information represents each worker’s overall capability such as processing power, network bandwidth, and the quality of service. Furthermore, this implicit information can be calculated at each coordinator node without the need to query other computing node for their current processing capacities. By combining our implicit information concept with the hierarchical load sharing structure, we propose new load sharing policy that can efficiently distribute workload among a large number of computing nodes while inflicting very small overhead. We use simulated experiments to compare our load sharing policy with other load sharing policies. The simulation results indicate that our proposed load sharing policy, which is simpler, can perform comparable to other load sharing policies, even when they have the perfect knowledge of computing powers when making load sharing decision.