Evaluating Cache Coherence in the DSMIO System

Abstract

There has been a significant amount of research on object-based database management systems (ODBMS). In contrast to traditional relational systems, ODBMS use the same data model for the clients and the server. Thus, parallel ODBMS typically adopt a data-shipping architecture, which allows data request processing to be performed at the clients. Although such a strategy can improve performance by moving data closer to clients and by alleviating the load on the server, it raises the issue of cache coherence. Our work is based on the observation that coherence maintenance techniques developed to improve softwarebased distributed shared-memory systems can be exploited to improve parallel ODBMS. In previous papers we proposed a coherence algorithm, called DSMIO, that uses the Lazy Release Consistency model and the diffing mechanism to limit the number and size of coherence messages and save cache space. These papers evaluated DSMIO in the context of a parallel object-based database system where one machine acted as a server, accessing stable storage and keeping client caches coherent. In this paper we evaluate DSMIO for a system where each machine can act both as a client and a server. We use the state-of-the-art CallBack Locking (CBL) cache coherence algorithm as a basis for comparison. Our results show that DSMIO significantly outperforms CBL for workloads with even a moderately high frequency of database updates. Keywords— Client-Server OODBMS, Parallel Database Systems, Cache Coherence, Distributed Shared Memory

5 Figures and Tables

Cite this paper

@inproceedings{Osthoff2000EvaluatingCC, title={Evaluating Cache Coherence in the DSMIO System}, author={Carla Osthoff and Cristiana Seidel and Ricardo Bianchini and Marta Mattoso and Claudio Luis de Amorim}, year={2000} }