A framework for intelligence augmented computing systems

Banerjee, Subho Sankar

Permalink

https://hdl.handle.net/2142/115749 Copy

Description

Title

A framework for intelligence augmented computing systems

Author(s)

Banerjee, Subho Sankar

Issue Date

2022-04-25

Director of Research (if dissertation) or Advisor (if thesis)

Iyer, Ravishankar K

Doctoral Committee Chair(s)

Iyer, Ravishankar K

Committee Member(s)

• Hwu, Wen-mei
• Adve, Vikram S
• Mitra, Subhasish

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Machine Learning for Systems
• Reinforcement learning
• Bayesian Methods
• Accelerators
• Scheduling
• Performance Monitoring
• Error Correction

Abstract

Large-scale computing systems rely on many control and decision-making algorithms. Classical approaches to designing and optimizing these algorithms are poorly suited to the diverse and demanding requirements of modern systems and emerging applications. The state of the art paradigm for building these control algorithms often devolves into painstakingly built, handcrafted, average-case heuristics. However, as systems and applications have grown in complexity and heterogeneity, designing fixed algorithms that work well across a variety of conditions has become exceedingly difficult and costly. Moreover, we are reaching the limits of conventional approaches of generating heuristics, which involve recurring human-expert-driven engineering efforts. Such an approach will be untenable in the future. In this thesis, we investigate a new paradigm for solving large scale system management and optimization problems. We develop systems that can learn to optimize the performance on their own using modern machine learning techniques. As a result, in the proposed approach, the system designer need not develop specialized heuristics for low-level design goals. Instead, the designer architects a framework for measurement, estimation, experimentation, and learning that discovers the low-level actions that achieve high-level resource management objectives automatically. We use this approach to build a series of practical intelligent controllers for the management and optimization of large-scale data-parallel and data-processing workloads on heterogeneous computer systems. Our contributions encompass building mathematical models (e.g., for denoising telemetry data), policies (e.g., for scheduling), optimizations to enable real time inference, and the design and implementation of practical software and hardware that provides efficient, scalable, and composable system management solutions.

Graduation Semester

2022-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Subho Banerjee

Owning Collections