Predicting Response Time of Concurrent Queries with Similarity Models

Abstract

Predicting query response time is an essential task for managing database systems, especially in modern large distributed data centers that execute heterogeneous query workloads concurrently. The core of such a model is to quantify query interaction, which is neglected by the state of the art models. This paper proposes a novel model that estimates query response time based on the similarity of query mixes. We introduce a notion called query rating for constructing the feature vector of a query and developed a measure of the similarity between two query mixes. We propose a static similarity model to estimate the response time of a query by using that of the most similar query mixes containing the query. We also build a dynamic model based on the static model to predict the remaining execution time of a query on-the-fly whenever a new query mix forms. A scheduling method is proposed with the similarity models as the key enablement, which schedules a workload with minimum execution time. The experimental evaluation shows that our models perform approximately 12% and 35% of the actual response time on average for static and dynamic respectively, and the scheduler with our model outperforms, up to 2.9x, that with conventional models consistently.