HPSF & KUG Presentation Ideas

[crtrott]

Please list as new comments:

• Title
• Short description (bullet points are fine)
• HPSF Track/Topic area if applicable

We want proposed talks and then we can further discuss & consolidate on what we want to put forward as a team.

[masterleinad]

Title:
Kokkos' approach to CI

Description:
In this presentation, we outline the Kokkos strategy for Continuous Integration (CI) and comprehensive software testing. The Kokkos CI framework is built on intentional redundancy, leveraging a diverse array of build systems and infrastructures. We will detail how this redundancy is distributed across different CI systems and facilities, while sharing our experiences and the specific pitfalls encountered with each.

Furthermore, we discuss the additional layers of testing that ensure long-term software robustness and reliability. These include:

• Nightly Testing: Expanding coverage across a wide spectrum of compilers, versions, configuration options, and architectures.
• Integration Testing: Regular synchronization with Trilinos—a scientific computing library that utilizes Kokkos as a foundational component—to provide stress testing outside the unit testing harness.

Topic/Track:
Productivity, Performance, and the HPC pipeline

Session format:
Session Presentation - 25 minutes

Audience Level:
Any

Benefits to the Ecosystem:
CI essential for all software projects and detailing how we build resilience in Kokkos
is beneficial for other projects

Presented this talk before:
No

[JBludau]

Title: Performance portable design patterns – A gallery and place to discuss them in the community

Session:
How to encourage collaboration

Description:
I present the project of my BSSw fellowship (if funded). It focuses on creating a public website for sharing, evolving, and curating software design patterns for heterogeneous computing.
Proven software design patterns that work in heterogeneous settings allow to transfer knowledge between domains, levels of experience, and institutions. Previous collections of design patterns like the “Design Patterns” book by the “Gang of 4” show, that providing a curated selection of recurring patterns and their implementation is very impactful for the community.
But in the age of heterogeneous computing, with its restrictions due to memory spaces and non-sequential parallelism as well as collective execution, new design patterns are required. But these patterns already exist within the HPC community. I will provide a definition of the design patterns of interest, show some examples, and present a website to collect and curate them. This is part of my project to facilitate the discussion, collection and evolution of portable design patterns in a community based sustainable way.

Relevance:
Domain scientists are not trained in software development but need to design code to work on today’s heterogeneous supercomputers. Additionally, heterogeneous computing restricts the software design thus slowing development and requiring refactoring or even redesign of scientific codes. Design patterns can help transfer knowledge within the community. Furthermore, the presented website serves as a place for discussion and evolution of these design patterns, fostering a sustainable knowledgebase in the community.

Track: How to encourage collaboration/HPC user experience

Format: 15mins

Title: Heterogeneous Memory Management – Blessing or Curse

Session: Future directions of HPC

Description:
Intel, Nvida and AMD all have long provided host-device migratable memory. But with the development of the Linux Heterogeneous Memory Management (HMM) kernel module, system allocated memory becomes migratable to the GPU, too.
This page-fault based mechanism of automatic memory movement allows to use memory in standard c++ containers like vector to be used on the GPU. While this might help porting code to the GPU, it has impacts on performance, synchronization behavior, and even the reliability of unit testing.
The proposed talk will explain what HMM is and which hardware/software combinations support it. Furthermore, it will discuss resulting code changes and highlight implications on synchronization behavior and unit testing.
Finally, the talk will dive into performance measurements comparing the performance of “classical”, disjunct memory management and HMM based management on a variety of architectures.
Relevance:
Vendors try to sell migrating memory as solution to memory management in heterogeneous computing. Nevertheless, developers should be aware of the limitations, restrictions and implications on their unit tests. Furthermore, writing a code assuming all memory to be migrating makes switching back hard. Meanwhile, writing code with manual memory management allows the code to work with and without HMM. This highlights which strategy to choose for a sustainable software development in a world with uncertain next-generation hardware.

Track: Future direction of HPC

Format: 25mins

Title: mdspan – A chance for interoperability

Session: HPC user experience

Description:
Mdspan was integrated into the c++ standard with c++-23. Thus, c++ finally has a (non-owning) multidimensional memory abstraction to represent the physical objects in many scientific applications. The performance portability library Kokkos uses a device compatible version of mdspan since its 5.0 release. It is expected that other libraries and applications will adapt mdspan in the future.
This provides a unique possibility to make interoperability with other languages like python easier. Python uses an abstraction named “buffer” that resembles mdspan. Therefore, providing an interoperability layer for mdspan <-> buffer would allow a large number of c++ based code to interoperate seamlessly with python. This is especially relevant in the age of AI/ML which is dominated by python while a lot of HPC software is written in c++.
This talk presents an implementation of an interoperability layer between c++’s mdspan and python’s buffer. The usefulness is demonstrated by showing the integration of a ML model into a c++ code. Furthermore, the talk lists the current limitations and next steps to make the implementation more complete.
Relevance:
Providing interoperability for foundational c++ types like mdspan allows a large set of applications to interoperate with various python packages via a common interface. This greatly reduces the complexity of bindings and wrappers and allows easy collaboration between AI/ML models and HPC software. Upstreaming the implementation helps making interoperability a “works out of the box” feature in the future, allowing users to easily combine libraries into new software.

Track: How to encourage collaboration/HPC user experience

Format: 15mins

[cwpearson]

Title: Navigating Hardware Evolution in the Kokkos Ecosystem

Description:
Novel HPC architectures demand that performance-oriented scientific software projects must adapt, but what does this process look like in practice? This talk explores hardware/software co-design efforts in the Kokkos ecosystem, examining the challenges, workflows, and collaborative strategies required to align Kokkos and emerging hardware capabilities. We discuss why performance portability layers like Kokkos enable productive dialogue between application developers and hardware vendors while insulating scientific codes from architectural churn. Rather than focusing on specific performance outcomes, we convey lessons and insights for other HPC software projects considering co-design efforts.

HPSF Track:
Emerging Technologies and the Future of HPC (15mins or 25mins)

This might also work as a 20 minute KUG talk, with the emphasis adjusted somewhat.

[diehlpk]

Title: Automatic Translation of Fortran to Kokkos

Description:

Tentative talking points:
(1) AST generation
(2) Preprocessing/DSL support
(3) Increasingly complex examples

HPSF Track:
Emerging Technologies and the Future of HPC (15mins or 25mins)