Multi-metric Graph Query Performance Prediction

We propose a general framework for predicting graph query performance with respect to three performance metrics: execution time, query answer quality, and memory consumption. The learning framework generates and makes use of informative statistics from data and query structure and employs a multi-label regression model to predict the multi-metric query performance. We apply the framework to study two common graph query classes—reachability and graph pattern matching; the two classes differ significantly in their query complexity. For both query classes, we develop suitable performance models and learning algorithms to predict the performance. We demonstrate the efficacy of our framework via experiments on real-world information and social networks. Furthermore, by leveraging the framework, we propose a novel workload optimization algorithm and show that it improves the efficiency of workload management by 54% on average.