Decades of research in the field of database management systems (DBMSs) have focused on improving system performance. Modern analytical systems leverage innovative execution methods, such as vectorization and compilation, or enable parallelizing execution at the operator level to reduce single-query runtimes. Unfortunately, further developments to improve single-query execution performance have failed to yield significant improvements and are providing diminishing performance returns.

To extend beyond the limits of single-query performance improvements, we propose a co-design method to align database and queueing theory research in workload and architecture-aware scheduling policies. In this work, we present the addition of a scheduling component to a highly optimized execution engine and the design of new scheduling algorithms combining awareness of query characteristics and the execution hardware. Our proposed scheduling policies order and assign query sub-tasks to compute resources to enhance performance on analytical workloads in a modern, in-memory execution environment. Further improvements to the execution framework address imbalanced data access patterns and enable locality-aware execution. By optimizing for resource efficiency, our developments decrease the average system latency by over 30%.

77 pages

Thesis Committee:
Andy Pavlo (Chair)
Justine Sherry

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by [email protected]