Advances in parallel programming for electronic design automation

Lin, Chun-Xun

Permalink

https://hdl.handle.net/2142/108425 Copy

Description

Title

Advances in parallel programming for electronic design automation

Author(s)

Lin, Chun-Xun

Issue Date

2020-07-07

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

Wong, Martin D F

Doctoral Committee Chair(s)

Wong, Martin D F

Committee Member(s)

• Hwu, Wen-Mei
• Chen, Deming
• Xiong, Jinjun

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Electronic design automation
• Parallel programming

Abstract

The continued miniaturization of the technology node increases not only the chip capacity but also the circuit design complexity. How does one efficiently design a chip with millions or billions transistors? This has become a challenging problem in the integrated circuit (IC) design industry, especially for the developers of electronic design automation (EDA) tools. To boost the performance of EDA tools, one promising direction is via parallel computing. In this dissertation, we explore different parallel computing approaches, from CPU to GPU to distributed computing, for EDA applications. Nowadays multi-core processors are prevalent from mobile devices to laptops to desktop, and it is natural for software developers to utilize the available cores to maximize the performance of their applications. Therefore, in this dissertation we first focus on multi-threaded programming. We begin by reviewing a C++ parallel programming library called Cpp-Taskflow. Cpp-Taskflow is designed to facilitate programming parallel applications, and has been successfully applied to an EDA timing analysis tool. We will demonstrate Cpp-Taskflow’s programming model and interface, software architecture and execution flow. Then, we improve Cpp-Taskflow in several aspects. First, we enhance Cpp-Taskflow’s usability through restructuring the software architecture. Second, we introduce task graph composition to support composability and modularity, which makes it easier for users to construct large and complex parallel patterns. Third, we add a new task type in Cpp-Taskflow to let users control the graph execution flow. This feature empowers the graph model with the ability to describe complex control flow. Aside from the above enhancements, we have designed a new scheduler to adaptively manage the threads based on available parallelism. The new scheduler uses a simple and effective strategy which can not only prevent resource from being underutilized, but also mitigate resource over-subscription. We have evaluated the new scheduler on both micro-benchmarks and a very-large-scale integration (VLSI) application, and the results show that the new scheduler can achieve good performance and is very energy-efficient. Next we study the applicability of heterogeneous computing, specifically the graphics processing unit (GPU), to EDA. We demonstrate how to use GPU to accelerate VLSI placement, and we show that GPU can bring substantial performance gain to VLSI placement. Finally, as the design size keeps increasing, a more scalable solution will be distributed computing. We introduce a distributed power grid analysis framework built on top of DtCraft. This framework allows users to flexibly partition the design and automatically deploy the computations across several machines. In addition, we propose a job scheduler that can efficiently utilize cluster resource to improve the framework’s performance.

Graduation Semester

2020-08

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/108425

Copyright and License Information

Copyright 2020 Chun-Xun Lin

Owning Collections