Customizable Computing: Automated Cross-Layer Optimizations for Efficiency and Security - Cyber

Abstract

As the semiconductor fabrication technology scales down to tens of nanometers, we are starting to reach physical limits that make ideal scaling practically impossible. As a result, future computing systems will need to improve their efficiency by closely matching hardware and software to real application needs rather than primarily relying on efficiency gains from the technology scaling. In other words, computing systems need to be customized or specialized to applications. In today’s computing systems, however, heavy customization comes with the significant costs in increased design and verification efforts. This project aims to address this challenge by developing a hardware-software framework that enables automated customization and optimization of computing systems across traditional abstraction layers. The automated customization will remove inefficiencies in the abstraction layers with minimal designer efforts and also improve security by introducing diversity. The project consists of two main technical directions: (i) hardware-assisted software customization and (ii) cross-abstraction optimizations for reducing dynamic inefficiencies. The project will investigate a number of run-time invariants for software customization, including unused code and constant variables. Our preliminary study suggests that there is a large potential room for improvements. Typical programs use less than a half of its code at run-time. We also found that over 30% of run-time instructions are used to eventually produce constant values in many cases. The project will develop hardware support that can enable aggressive customizations with low overhead while providing correctness and real-time guarantees. For the second thrust to reduce run-time inefficiencies, the project will investigate time-driven resource allocation for interactive applications, and memory hierarchy optimizations for accelerators. The preliminary study suggests that an ideal predictor that can adjust voltage and frequency levels at a fine-granularity based on the response time target has a potential to reduce the energy consumption by over 80% compared to today’s conservative approach to meet the worst-case performance need all the time. The capability to automatically customize a computing system benefit defense computing systems by improving the efficiency and the security. While the proposed techniques are applicable to computing systems in general, the techniques will be particularly well-suited for defense systems that have well-define missions and tightly controlled software.

Document Details

Document Type
DoD Grant Award
Publication Date
Aug 08, 2016
Source ID
N000141512175

Entities

People

  • Edward Suh

Organizations

  • Cornell University
  • Office of Naval Research
  • United States Navy

Tags

Fields of Study

  • Computer science

Readers

  • Distributed Systems and Data Platform Development
  • Parallel and Distributed Computing.

Technology Areas

  • Cyber
  • Microelectronics