OpenMP parallelism in scientific computing

Abstract:

The purpose of the course is to learn when a code could benefit from task-based parallelism, and how to apply it. A task-based algorithm comprises of a set of self-contained tasks that have well-defined inputs and outputs. This differs from the common practice of organizing an implementation into subroutines in that a task-based implementation does not call the associated computation kernels directly, instead it is the role of a runtime system to schedule the task to various computational resources, such as CPU cores and GPUs. One of the main benefits of this approach is that the underlying parallelism is exposed automatically as the runtime system gradually traverses the resulting task graph.

Content:

The course mainly focuses on the task-pragmas implemented in the newer incarnations of OpenMP. Other task-based runtime systems, e.g., StarPU, and GPU offloading are briefly discussed.

Format:

The course will be three half-days and comprises of lectures and hands-on sessions. This is an online-only course (Zoom).

Audience:

…

Date and Time:

2021-05-{10,11,12}, 9:00-12:00

Location:

Online through Zoom

Instructors:

Pedro Ojeda-May (pedro.ojeda-may@umu.se)

Original author:

Mirko Myllykoski.

Recordings:

…

Registration:

…

Acknowledgment:

This material was orignally created by Mirko Myllykoski (mirkom@cs.umu.se), and later modified.

Day 1

• Introduction to HPC2N
• Introduction to Kebnekaise
• Introduction to OpenMP (part 1)
• Introduction to OpenMP (part 2)

Day 2

• Introduction to task-based parallelism
• OpenMP task basics (part 1)
• OpenMP task basics (part 2)
• Example: LU factorization

Day 3

• Demonstration: Two-sided matrix algorithms
• StarPU runtime system (part 1)
• StarPU runtime system (part 2)
• Example: LU factorization

Miscellanous

• Quick Reference
• Instructor’s guide