Building blocks for high-performance languages

Szaday, Justin Josef

Permalink

https://hdl.handle.net/2142/116159 Copy

Description

Title

Building blocks for high-performance languages

Author(s)

Szaday, Justin Josef

Issue Date

2022-07-06

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

Kale, Laxmikant

Doctoral Committee Chair(s)

Kale, Laxmikant

Committee Member(s)

• Amato, Nancy
• Padua, David
• Chamberlain, Brad

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)

• Compilers
• Programming Languages
• High-performance Computing
• Runtime Adaptivity

Abstract

As machine complexity and diversity rise, the importance of high-level abstractions to the high-performance computing ecosystem rises. These tools, like adaptive runtimes and domain-specific or general-purpose languages, promote developer productivity and performance portability, lowering the barrier to building large-scale, high-performance parallel applications. However, their development and maintenance costs can be discouragingly high, relegating us to the use of classical, low-level tools. Observing this, we propose that a language framework with high-level constructs built on a rich runtime system can significantly improve the productivity of language authors. To explore this hypothesis, we designed and implemented such a framework, EIR, then targeted it with an extensible language, Ergoline. Within Ergoline, we embedded several DSLs derived from past works built on the underlying RTS, Charm++. We explored code reuse among them and offered shared optimizations within EIR. To close semantic gaps between Charm++ and EIR, we developed Hypercomm—a suite of optimized abstractions. Through our experiences developing these tools, we conclude that our approach can offer performance and productivity benefits to language authors and their users alike

Graduation Semester

2022-08

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Justin Szaday

Owning Collections