Performance Evaluation of Parallel Transaction Processing in Shared Nothing Database Systems

Abstract. Complex and data-intensive database queries mandate parallel processing strategies to achieve sufficiently short response times. In praxis, parallel database processing is mostly based on so-called "shared nothing " architectures entailing a physical partitioning and alloca-tion of the database among multiple processing nodes. We examine the performance of such architectures by using a detailed simulation system. We analyse response time performance of transactions and individual database queries in single-user as well as in multi-user mode. Fur-thermore, we study the throughput behavior for on-line transactions. Three workload types covering a wide range of commercial applications are used for performance evaluation: the de-bit-credit benchmark load, synthetically generated relational queries as well as real-life work-loads represented by database traces.