ANALYSIS OF BACKEND DATABASE SYSTEMS ARCHITECTURES

The concept of offloading the data management function to a dedicated processor has been suggested many times as a quick and economical method for improving the performance and resource utilization of a data processing installation with a heavy database workload. A backend database machine becomes more cost effective when it is constructed using a number of cheap mini or microprocessors. In this paper we present two multiprocessor backend database architectures and investigate their performance using queueing network analysis methods. The first architecture uses multiple processors in a pipelined fashion to construct an intelligent disk controller. The second employs multiple loosely coupled backend processors each with its own disk. We specifically do not consider any special purpose hardware devices because it is interesting to see how much performance gain can be achieved by using just the existing components. The models are parameterized based upon information available from previous studies and descriptions of state-of-the art hardware. An on-line transaction processing workload is assumed. The analysis indicates the potential superiority of the multiple backends architecture over the pipelined controller architecture. The methodology presented here can be extended for other workloads and configurations. In addition to the performance projections, the paper provides analysis of the impact of the various backend configurations upon some of the basic objectives of data management systems such as robustness, security and reduction of application complexity.