Partitioning vs. Replication for Token-Based Commodity
Franklin, Michael J.
MetadataShow full item record
The proliferation of e-commerce has enabled a new set of applications that allow globally distributed purchasing of commodities such as books, CDs, travel tickets, etc., over the Internet. These commodities can be represented on line by tokens, which can be distributed among servers to enhance the performance and availability of such applications. There are two main approaches for distributing such tokens ? replication and partitioning. Token replication requires expensive distributed synchronization protocols to provide data consistency, and is subject to both high latency and blocking in case of network partitions. On the other hand, token partitioning allows many transactions to execute locally without any global synchronization, which results in low latency and immunity against network partitions. In this paper, we examine the Data-Value Partitioning (DVP) approach to token-based commodity distribution. We propose novel DVP strategies that vary in the way they redistribute tokens among the servers of the system. Using a detailed simulation model and real Internet message traces, we investigate the performance of our DVP strategies by comparing them against a previously proposed scheme, Generalized Site Escrow (GSE), which is based on replication and escrow transactions. Our experiments demonstrate that, for the types of applications and environment we address, replication-based approaches are neither necessary nor desirable, as they inherently require quorum synchronization to maintain consistency. We show that DVP, primarily due to its ability to provide high server autonomy, performs favorably in all cases studied. (Also cross-referenced as UMIACS-TR-2000-62