Adaptive Concurrent Query Execution Framework for an Analytical In-Memory Database System

Abstract

There is a growing need for in-memory database analytic services, especially in cloud settings. Concurrent query execution is common in such environments. A crucial deployment requirement is to employ a concurrent query execution scheduling framework that is flexible, precise, and adaptive to meet specified deployment goals. In addition, the framework must also aim to use all the underlying hardware resources effectively (for high performance and high cost efficiency). This paper focuses on the design and evaluation of such a scheduler framework. Our scheduler framework incorporates a design in which the scheduling policies are cleanly separated from the scheduling mechanisms, allowing the scheduler to support a variety of policies, such as fair and priority scheduling. The scheduler also contains a novel learning component to monitor and quickly adapt to changing resource requirements of concurrent queries. In addition, the scheduler easily incorporates a load controller to protect the system from thrashing in situations when resources are scarce/oversubscribed. We have implemented our scheduling framework in an in-memory database engine, and using this implementation we also demonstrate the effectiveness of our approach. Collectively, we present the design and implementation of a scheduling framework for in-memory database services on contemporary hardware in modern deployment settings.