On the performance of concurrent tree algorithms

Various algorithms have been proposed to allow independent concurrent processes to access shared tree-structured objects, while guaranteeing that the object remains in a consistent state. Although algorithms can be classified as either “optimistic” or “conservative,” on the basis of their approach to concurrency control, it is very difficult to compare the expected performance of two concurrent algorithms that use different approaches. We demonstrate an approach to performance comparison that combines both analytical evaluation and detailed simulation. Our analysis predicts the frequency of highly time dependent activities, such as interference and recovery, that dominate performance of optimistic algorithms. Our simulation measures the performance of concurrent algorithms in a particular application, that of a central file server. The simulation results show that in general analytical techniques tend to overestimate the performance of conservative algorithms because inherent system bottlenecks are ignored, whereas they underestimate the performance of optimistic algorithms due to the difficulty of obtaining tight analytical bounds on time-dependent interactions. However, the results also show that our form of analysis gives predictions that much more closely match simulation results.