Dynamic load balancing in RCAN content addressable network

Abstract

RCAN [1] is a novel multi-ring content addressable peer-to-peer system. RCAN was proposed in the aim of improving the routing performance of CAN [4] overlays while minimizing the maintenance overhead during nodes churn in large networks. The key idea of RCAN is to equip each node with a few long-links towards some distant nodes. Long-links are clockwise directed and wrap around to form small rings along each dimension. The number of rings and their sizes self-adjust as nodes join and leave the system. RCAN is a pure P2P design, where all nodes assume the same responsibility. Unlike some existing P2P overlays, RCAN is self-organizing and does not assume any a-priori fixed limits for the network size or the routing state per node. Each node auto-adapts its routing state to cope with network changes.In this work, we address the problem of load balancing in RCAN system. We propose fully decentralized mechanisms to cope with large scale data imbalance. Our methods extend the random join process [7,12] to enable a newly joining node to locate an existing node with a heavy load to share the load with. A new joining node selects one or more existing nodes at random, and then it chooses the heaviest node among them to share the load with. The heaviest node may also be located through a random walk starting from a randomly selected node. In opposition to [7], the proposed methods do not rely on any external index nor assume any global knowledge. All routing and load information is disseminated in the network through links established between immediate and distant neighbors, incurring thus no extra overhead. On another side, the number of probed nodes is logarithmic [1] which enables to achieve constant load imbalance less than or equal to 8 [12, 13, 16]. A combination of the two methods would achieve much smaller load imbalance that is less than or equal to 4 [12]. We also studied the impact of number of entry points to the systems on the load imbalance factor. To the best of our knowledge we are the first to address this problem. We have also conducted an extensive experimental study using uniform and non-uniform data distributions to demonstrate the effectiveness and the scalability of our proposals.