Distributed offline load balancing in MapReduce networks

In this paper we address the problem of balancing the processing load of MapReduce tasks running on heterogeneous clusters, i.e., clusters composed of nodes with different capacities and update cycles. We present a fully decentralized algorithm, based on ratio consensus, where each mapper decides the amount of workload data to handle for a single user job using only job specific local information, i.e., information that can be collected from directly connected neighboring mappers, regarding their current workload usage and capacity. In contrast to other algorithms in the literature, the proposed algorithm can be deployed in heterogeneous clusters and can operate asynchronously in both directed and undirected communication topologies. The performance of the proposed algorithm is demonstrated via simulation experiments on large-scale strongly connected topologies.