Towards predicting query execution time for concurrent and dynamic database workloads

摘要: Predicting query execution time is crucial for many database management tasks including admission control, scheduling, and progress monitoring. While a number of recent papers have explored this problem, the bulk existing work either considers prediction single query, or static workload concurrent queries, where by "static" we mean that queries to be run are fixed known. In paper, consider more general problem dynamic workloads. Unlike most previous on prediction, our proposed framework based analytic modeling rather than machine learning. We first use optimizer's cost model estimate I/O CPU requirements each pipeline in isolation, then combination queueing buffer pool merges requests from predict running times. compare approach with machine-learning variant work. Our experiments show analytic-model can lead competitive often better accuracy its counterpart.