Availability Benchmarking of a Database System

We present the results of an availability benchmarking study of a three-tier transactionprocessing-oriented database system based on Microsoft SQL Server 2000. Following the general availability benchmarking methodology introduced by our previous work on software RAID systems, we carried out a set of fault-injection experiments in which we measured the effects of 14 different types of realistic storage system faults on the performance and success rate of database transactions drawn from a TPC-C-like OLTP workload. Our results indicate that the SQL Server 2000 database is robust to a wide range of storage faults, including most single-sector uncorrectable I/O errors, although it does not transparently tolerate entire disk failures or timeouts. We also identify transactions and writeahead logging as important mechanisms used to achieve this level of robustness, and furthermore identify the system’s middleware component as a key weakness in obtaining a high level of end-to-end fault tolerance.