Introducing principled approximation and online control into streaming applications




Pei, Yan, Ph. D.

Journal Title

Journal ISSN

Volume Title



The ubiquity of streaming applications in important domains such as deep learning, computer vision/graphics, Internet of Things has opened up opportunities for the use of approximate computing to enable efficient execution of these applications on a wide range of platforms. This dissertation explores the use of ideas from machine learning and control theory to exploit approximation in streaming applications in a principled way. We first present online control techniques of introducing principled approximation into Simultaneous Localization and Mapping (SLAM) algorithms, which are used in emerging domains like robotics and autonomous driving in which autonomous agents build a map while navigating through unknown environments. Existing studies of approximation in SLAM have mostly used offline control, assuming the trajectory is known before the agent starts to move, which is impractical. The proposed methodology controls approximation in an adaptive manner without causing unacceptable quality degradation, enabling efficient deployment of SLAM on a wider range of resource-constrained platforms. We also propose Sonic, a sampling-based online controller for general constrained optimization problems in long-running streaming applications. Within each phase of a streaming application’s execution, Sonic utilizes the beginning portion to sample the knob space sequentially and aims to pick the optimal knob setting for the rest of the phase, given a user-specified constrained optimization problem. Machine learning regressors and Bayesian optimization are applied in Sonic for better sampling choices. Our experiments show that Sonic is able to find near-optimal knob settings at run time for the applications we studied. For online control, state estimation is a key component. In this dissertation, we introduce a novel derivation of Kalman filtering, a classic state estimation technique that can be used in online control to combine noisy estimates of quantity of interest. It is presented from an abstract perspective with key assumptions and concepts clarified. We then exploit these insights and propose RASR, a performance-oriented super-resolution program for rendered content that takes advantage of internal sub-pixel states of graphics hardware. RASR is a deep learning approach to fusing frames of low-fidelity into one highfidelity frame, designed to meet the increasing demand for high throughput, high quality and low latency in real-time rendering.


LCSH Subject Headings