| Summary: | The typical workload in a database system consists of a mix of multiple queries of different types, running concurrently and
interacting with each other. The same query may have different performance in different mixes. Hence, optimizing performance
requires reasoning about query mixes and their interactions, rather than considering individual queries or query types. In this
dissertation, we demonstrate how queries affect each other when they are executing concurrently in different mixes. We show the
significant impact that query interactions can have on the end-to-end workload performance. A major hurdle in the understanding of query interactions in
database systems is that there is a large spectrum of possible causes of interactions. For example, query interactions can happen
because of any of the resource-related, data-related or configuration-related dependencies that exist in the system. This
variation in underlying causes makes it very difficult to come up with robust analytical performance models to capture and model query
interactions. We present a new approach for modeling performance in the presence of interactions, based on conducting experiments to measure the effect of query interactions and fitting statistical models to the data collected in these experiments to capture the impact of query interactions. The experiments collect samples of the different possible query mixes, and measure the performance metrics of interest for the different queries in these sample mixes.
Statistical models such as simple regression and instance-based learning techniques are used to train models from these sample mixes. This approach requires no prior assumptions about the internal workings of the database system or the nature or cause of
the interactions, making it portable across systems. We demonstrate the potential of capturing, modeling, and exploiting query interactions by developing techniques to help in two database performance related tasks: workload scheduling and estimating the
completion time of a workload. These are important workload management problems that database administrators have to deal with
routinely. We consider the problem of scheduling a workload of
report-generation queries. Our scheduling algorithms employ statistical performance models to schedule appropriate query mixes
for the given workload. Our experimental evaluation demonstrates that our interaction-aware scheduling algorithms outperform scheduling policies that are typically used in database systems. The problem of estimating the completion time of a workload is an important problem, and the state of the art does not offer any systematic solution. Typically database administrators rely on heuristics or observations of past behavior to solve this problem. We propose a more rigorous solution to this problem, based on a workload simulator that employs performance models to simulate the execution of the different mixes that make up a workload. This mix-based simulator provides a systematic tool that can help database administrators in estimating workload completion time. Our
experimental evaluation shows that our approach can estimate the workload completion times with a high degree of accuracy. Overall, this dissertation demonstrates that reasoning about query
interactions holds significant potential for realizing performance improvements in database systems. The techniques developed in this work can be viewed as initial steps in this interesting area of research, with lots of potential for future work.
|
|---|
