This research is concerned with making atomic transactions perform well in the wide-area distributed systems that will be the norm in the future. The work focuses on support for consistency levels that are weaker than serializability; these levels are of interest because many applications do not require full serializability and can achieve better performance by running at weaker levels. Some consistency levels have been defined in the past, but the earlier definitions are inadequate because they either do not rule out undesirable behaviors, or they limit implementation techniques unnecessarily. In particular they do not allow optimistic techniques, which is unfortunate because these are the best for wide area and mobile systems. This research provides definitions of existing levels that are both constrained enough to rule out bad behaviors and permissive enough to allow a wide range of concurrency control mechanisms. In addition the research defines a number of useful new levels. Furthermore, the research develops optimistic implementations of the weak levels, and uses simulation and implementation to investigate the performance advantages that can be achieved by running at weaker levels. The research will lead to precise definitions of new and old weak levels, to a set of new, optimistic implementations of the levels, and to a thorough understanding of the performance benefits that derive from weaker levels. www.pmg.lcs.mit.edu
