Conflict-minimizing dynamic load balancing for P2P desktop Grid

Abstract

Fully decentralized resource allocation for P2P desktop Grid allows each participating node to act as both resource provider and requester. The system performance indicators (including throughput, makespan, etc) are easily degraded by the unbalanced load distribution, which is probably caused by the fast-changing states of heterogeneous resources due to arbitrary task submissions. Although the cooperative load rebalancing methods can mitigate the problem, they are likely to introduce the contention on under-utilized resources with growing task arrival rates, leading to the sub-optimal load balancing efficacy. Our focus is on how to optimize load balancing status by taking into account minimizing the conflict of autonomic task migration decisions in P2P desktop Grid. Our load rebalancing process is modeled as a set of independent stochastic Bernoulli trials by letting each heavily loaded node push its surplus loads to its surrounding lightly loaded nodes. We proved that the surplus load amount should be shifted based on a proper ratio by considering decision conflicts and designed a novel load balancing algorithm with provably small decision conflict probability. We derived an upper-bound for this probability, which will be reduced down to about 2% under our algorithm. Finally, we validated via simulation that the system performance can be significantly improved accordingly.