Distributed dense linear algebra operations with Charm++

Array abstractions to represent linear algebra kernels and the growing popularity of Python for Data Science have increased interest in providing support for array types as first-class citizens. Similarly, scientific codes heavily rely on linear algebra kernels which are written primarily in C/C++ and Fortran with frameworks like Message Passing Interface (MPI) and OpenMP. These presents two extremes of the modern day linear algebra kernel implementations. Array abstractions while providing readability, fails to provide distributed scalability; and scientific codes implementing linear algebra while being performant, fails to be flexible and interoperable. This thesis introduces LibCharmTyles, a C++ library on top of Charm++ that supports array types as first-class citizens. Furthermore, LibCharmTyles supports both shared and distributed memory parallelism and provides linear scaling through over-decomposition, a key asset of Charm++. LibCharmTyles is conformant to Basic Linear Algebra Subprograms (BLAS) operations and supports scalar, vector, and matrix types. The thesis goes over the design of LibCharmTyles, and then explores its performance relative to single-node NumPy and cuNumeric, a drop-in replacement for NumPy supporting distributed memory.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Gupta, Nikunj

Contributors dc:contributor

• Kale, Laxmikant V.

• Charm++
• Parallel And Distributed Computing
• Linear Algebra

Statement dc:rights

• Copyright 2022 Nikunj Gupta

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/117684

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F117684