Code generation of array constructs for distributed memory systems

Programming for high-performance systems to fully utilize the potential of the computing system is a complex problem. This is particularly evident when programming distributed memory clusters containing multiple NUMA chips and GPUs on each node since it would require a complex combination of MPI, OpenMP, CUDA, OpenCL, etc to achieve high performance even for sequentially simplistic codes. Programs requiring high performance are usually painstakingly written by hand in C/C++ or Fortran using MPI+X to target these machines. This work presents a multi-layer code generation framework Vaani that takes a very high-level representation of computations and generates C+MPI code by transforming the input through a series of intermediate representations. The very high-level nature of the language greatly facilitates programming parallel systems. Additionally, the use of multiple representations provide a flexible and transparent venue for the user to interact and customize the transformation process to generate code suitable to the user and the target machine. Experimental evaluation shows that the current implementation of Vaani generates code that is competitive with handwritten codes and hand-optimized libraries.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2021

Author dc:creator

• Pothukuchi, Sweta Yamini

Contributors dc:contributor

• Padua, David
• Kale, Laxmikant V
• Kloeckner, Andreas
• Franchetti, Franz

• Code Generation
• Distributed Memory Sytems
• MPI
• High Level Abstractions

Statement dc:rights

• Copyright 2020 Sweta Yamini Pothukuchi

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/109402

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/109402