摘要

In this paper, we revisit the problem of load-balancing structured peer-to-peer systems with on-line protocols. Load-balancing is of major significance for large-scale decentralized networks in terms of enhanced scalability and performance. The main incentives behind balancing schemes are under-utilization of bandwidth and computer resources. Therefore, our methods focus mainly on task-skew. Specifically, we address the problem with on-line protocols on the basis of migration and enhanced availability. In particular, the cornerstones of our methods are the notions of virtual nodes, replication and Multiple realities, combined altogether with allocation techniques based on balls-in-bins games. The rationale of our dynamic protocol to depend exclusively on peer load distribution preserves intact the structural properties and search efficiency of the overlay used as an indexing infrastructure, while preserving the semantic information of the data (e.g., range partitioned network). We also propose an effective load-aware mechanism to facilitate robust operations that counteract against contingent churn failures. Finally, our work is complemented with extensive experiments using both real and synthetic data sets.

  • 出版日期2012-9

全文