Performance Characteristics of Epsilon Serializability with Hierarchical Inconsistency Boundsy

Epsilon serializability (ESR) is a weaker form of correctness designed to provide more concurrency than classic serializability (SR) by allowing, for example, query transactions to view inconsistent data in a controlled fashion i:e: limiting the inconsistency within the speciied bounds. In the previous literature on ESR, inconsistency bounds have been speciied with respect to transactions or with respect to objects. In this paper, we introduce the notion of hierarchical inconsistency bounds that allows inconsistency to be speciied at diierent granularities. The motivation for this comes from the way data is usually organized, in hierarchical groups, based on some common features and interrelationships. Bounds on transactions are speciied at the top of the hierarchy, while bounds on the objects are speciied at the bottom and on groups in between. We also discuss mechanisms needed to control the inconsistency so that it lies within the speciied bounds. While executing a transaction, the system checks for possible violation of inconsistency bounds bottom up, starting with the object level and ending with the transaction level. Thus far, to our knowledge, no work has been done to determine the quantitative performance improvement resulting from ESR. Hence in this paper we report on an evaluation of the performance improvement due to ESR incorporating hierarchical inconsistency bounds. The tests were performed on a prototype transaction processing system that uses timestamp based concurrency control. For simplicity, our implementation uses a two level hierarchy for inconsistency specii-cation-the transaction level and the object level. We present the results of our performance tests and discuss how the behavior of the system is innuenced by the transaction and object level inconsistency bounds. We make two important observations from the tests. First, the thrashing point shifts to a higher multiprogramming level when transaction inconsistency bounds are increased. Further, for a particular multiprogramming level and a particular transaction inconsistency bound, the throughput does not increase with increasing object inconsistency bounds but peaks at some intermediate value.