Task Scheduling for Highly Concurrent Analytical and Transactional Main-Memory Workloads

Task scheduling typically employs a worker thread per hardware context to process a dynamically changing set of tasks. It is an appealing solution to exploit modern multi-core processors, as it eases parallelization and avoids unnecessary context switches and their associated costs. Naively bundling DBMS operations into tasks, however, can result in sub-optimal usage of CPU resources: highly contending transactional workloads involve blocking tasks. Moreover, analytical queries assume they can use all available resources while issuing tasks, resulting in an excessive number of tasks and an unnecessary associated scheduling overhead. In this paper, we show how to overcome these problems and exploit the performance benefits of task scheduling for main-memory DBMS. Firstly, we use application knowledge about blocking tasks to dynamically adapt the number of workers and aid the OS scheduler to saturate CPU resources. In addition, we show that analytical queries should issue a low number of tasks in cases of high concurrency, to avoid excessive synchronization, communication and scheduling costs. To achieve that, we maintain a concurrency hint, reflecting recent CPU availability, that partitionable analytical operations can use as a limit while adjusting their task granularity. We integrate our scheduler into a commercial main-memory column-store, and show that it improves the performance of mixed workloads, by up to 12.5% for analytical queries and 370% for transactional queries.