Sensitivity evaluation of dynamic load sharing in distributed systems

Load sharing improves performance of distributed systems by distributing load from heavilyloaded nodes to lightly-loaded nodes in the system. We consider two basic dynamic load sharing policies: sender-initiated and receiver-initiated. In the sender-initiated policy, a heavily-loaded node attempts to transfer work to a lightly-loaded node and in the receiver-initiated policy a lightlyloaded node attempts to get work from a heavily-loaded node. In most previous studies, the firstcome/first-served node scheduling policy has been used. In addition, analysis and simulations in these studies have been done under the assumption that service times and inter-arrival times of jobs are exponentially distributed. The behaviour of these policies is not clear when these assumptions are relaxed. We report the sensitivity of the performance of the sender-initiated and receiverinitiated policies to node scheduling policy, variance in service times, and variance in inter-arrival times. Our objective is to provide an intuitive understanding of the behaviour of these policies that transcends specific system and workload models and parameter values used.