Efficient Partial Replication to Improve Locality in Large-Scale Systems

This work addresses data replication and consistency maintenance in large scale systems. Based on the observation that partial replication of data allows to improve network traffic and to preserve locality, particularly in large scale systems, it focuses on the constraints necessary to maintain consistency when partial replication is used. More precisely, it points out that the transfer of information necessary to manage consistency requirements might cancel the decrease in network load and the locality properties offered by partial replication. It is shown that causal consistency criterion suffers this drawback, but, on the contrary, PRAM consistency criterion allows efficient partial replication. The usefulness of PRAM is illustrated with the BellmanFord shortest path algorithm, which takes profit of graph locality to partially replicate the variables and is solvable in PRAM distributed shared memories.