Document Type

Technical Report

Publication Date


Technical Report Number



Replication has been studied as a method of increasing the availability of a data item in a distributed database subject to component failures and consequent partitioning. The potential for partitioning requires that a protocol be employed which guarantees that any access to a data item is aware of the most recent update to that data item. By minimizing the number of access requests denied due to this constraint, we maximize availability. In the event that all access requests are reads, placing one copy of the data item at each site clearly leads to maximum availability. The other extreme, all access requests are write requests or are treated as such, has been studied extensively in the literature. In this paper we investigate the performance of systems with both read and write requests. We describe a distributed on-line algorithm for determining the optimal parameters, or optimal quorum assignments, for a commonly studied protocol, the quorum consensus protocol[9]. We also show how to incorporate these optimization techniques into a dynamic quorum reassignment protocol. In addition, we demonstrate via simulation both the value of this algorithm and the effect of various read-write rations on availability. This simulation, on 101 sites and up to 5050 links(fully- connected), demonstrates that the techniques described here can greatly increase data availability, and that the best quorum assignments are frequently realized at the extreme values of the quorum parameters.