Date of Award

August 2020

Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Qinru Qiu


concurrent kernel execution, GPU computing, reinforcement learning, task scheduling

Subject Categories



General purpose graphics processing units have become a computing workhorse for a variety of data- and compute-intensive applications, from large supercomputing systems for massive data analytics to small, mobile embedded devices for autonomous vehicles. Making effective and efficient use of these processors traditionally relies on extensive programmer expertise to design and develop kernel methods which simultaneously trade off task decomposition and resource exploitation. Often, new architecture designs force code refinements in order to continue to achieve optimal performance. At the same time, not all applications require full utilization of the system to achieve that optimal performance. In this case, the increased capability of new architectures introduces an ever-widening gap between the level of resources necessary for optimal performance and the level necessary to maintain system efficiency.

The ability to schedule and execute multiple independent tasks on a GPU, known generally as concurrent kernel execution, enables application programmers and system developers to balance application performance and system efficiency. Various approaches to develop both coarse- and fine-grained scheduling mechanisms to achieve a high degree of resource utilization and improved application performance have been studied. Most of these works focus on mechanisms for the management of compute resources, while a small percentage consider the data transfer channels. In this dissertation, we propose a pragmatic approach to scheduling and managing both types of resources – data transfer and compute – that is transparent to an application programmer and capable of providing near-optimal system performance.

Furthermore, the approaches described herein rely on reinforcement learning methods, which enable the scheduling solutions to be flexible to a variety of factors, such as transient application behaviors, changing system designs, and tunable objective functions. Finally, we describe a framework for the practical implementation of learned scheduling policies to achieve high resource utilization and efficient system performance.


Open Access

Included in

Engineering Commons