Issue No. 05 - May (2009 vol. 20)
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/TPDS.2008.145
Tim Moors , University of New South Wales, Sydney
Aaron Harwood , University of Melbourne, Melbourne
John Risson , University of New South Wales, Rowville
Many distributed hash tables (DHTs) resolve lookups in O(\log n) hops, where n is the number of nodes. One-hop DHTs give lower lookup latencies and lower lookup failure rates. However, it is hard to maintain large, wide-area one-hop topologies. We contribute aecast, a new topology dissemination algorithm for one-hop DHTs. It avoids expensive repair mechanisms and critical points of failure in existing one-hop DHTs. When a node discovers by anti-entropy that it has missed a topology update, it initiates "controlled flooding,” sending the update to nodes in the multicast tree that also missed the update. We compare aecast with a widely cited epidemic multicasting algorithm, pbcast, by analysis and simulation. Aecast gives at least fivefold fewer out-of-date nodes on average within one round of a topology update. We support it with a fault-tolerant topology agreement protocol, so that only legitimate topology changes propagate throughout the overlay. Consequently, we argue that one-hop DHTs deserve greater attention for Internet applications in which reasonably reliable nodes carry high lookup loads.
Distributed hash tables, epidemics, gossip, anti-entropy, multicast, reliability, wide-area networks, de Bruijn graph.
Tim Moors, Aaron Harwood, John Risson, "Topology Dissemination for Reliable One-Hop Distributed Hash Tables", IEEE Transactions on Parallel & Distributed Systems, vol. 20, no. , pp. 680-694, May 2009, doi:10.1109/TPDS.2008.145