作者: Chunqiang Tang , D Chen , Sandhya Dwarkadas , Michael L Scott , None
DOI: 10.1109/IPDPS.2004.1302942
关键词:
摘要: Summary form only given. Most distributed applications require, at least conceptually, some sort of shared state: information that is nonstatic but mostly read, and needed more than one site. At the same time, RPC-based systems such as Sun RPC, Java RMl, CORBA, .NET have become de facto standards by which communicate. As a result, state tends to be implemented either through redundant transmission deep-copy RPC parameters or ad-hoc, application-specific caching coherence protocols. The former option can waste large amounts bandwidth; latter significantly complicates program design maintenance. To overcome these problems, we propose middleware system works seamlessly with systems, providing them global, persistent store accessed using ordinary reads writes. Relaxed models aggressive protocol optimizations reduce bandwidth required maintain state. Integrated support for transactions allows chain calls update atomically. We focus on implementation challenges involved in combining transactions. In particular, describe transaction metadata table processes inside share data invisible other exchange modifications efficiently. Using microbenchmarks large-scale datamining application, demonstrate how integration transactions, facilitates rapid development robust, maintainable code.