Traditional mirrored disk systems provide high reliability by multiplexing disks. Performance is improved with parallel reads and shorter read seeks. However, writes must be performed by both disks, limiting performance.

A methodology for benchmarking distributed database management systems is proposed. A distributed environment is characterized in terms of the communication costs incurred in data movement between sites, the number of nodes that participate in processing a query and the data distribution scheme used in the network. These three major characteristics form a basis for eight query types that capture the query performance characteristics in the network. It is demonstrated that the performance characteristics of any distributed database management system can be captured by running queries based on these query types. Preliminary results obtained in applying the methodology in a single-user LAN environment are presented.>

With recent declines in the cost of semiconductor memory and the increasing need for high performance I/O disk systems, it makes sense to consider the design of large caches. In this paper, we consider the effect of caching writes. We show that cache sizes in the range of a few percent allow writes to be performed at negligible or no cost and independently of locality considerations.

A study is made of a known implementation of an optimistic concurrency-control algorithm for centralized database systems and improvements are suggested to the algorithm. The authors propose an implementation of an algorithm for the basic timestamp-ordering concurrency control in centralized database systems. The two algorithms are compared by simulation experiments. As expected, the optimistic approach is better for transaction mixes dominated by retrievals. For transaction mixes dominated by updates, the optimistic algorithm spends time performing operations that have a good chance of being voided by earlier conflicting operations. The authors expected the timestamp algorithm to be better in these circumstances but, this is not the case. They attribute this to the fact that transactions used in the experiments are short, and therefore the execution time lost due to an abort is small.>