2021 ECP Community BOF Days - Exascale Computing Project

March 30, 11:00 am - The Tiered Testing Approach to Software Quality Assurance at Exascale and Beyond

Abstract:

A substantial portion of ECP’s efforts is the creation of sustainable processes to improve software quality today and well into the future. A key factor in this is testing. In this session, we will discuss ECP’s goals for creating ecosystems amenable to improved testing at any scale: from a developer’s workstation to the largest Exascale systems and beyond. This tiered approach to testing encompasses many different technologies and techniques. In particular, we’ll discuss ongoing CI efforts at the facilities, preparing validation tests for E4S, and early usage of the Spack testing features.

Speakers:

Tim Haines (University of Wisconsin-Madison)
Ryan Adamson (ORNL)
Jim Willenbring (Sandia)

March 30, 11:00 am - Exascale Computing Project Data Management, Data Analytics and Visualization Overview

Slides from the Event:

Abstract:

This Community BOF will provide an overview of data management and Data Analysis & Visualization capabilities under development as part of the Exascale Computing Project (ECP). We will cover the data management portfolio and the Data Analysis and Visualization portfolio.

Changes in the hardware architecture of Exascale supercomputers will render current approaches to data management, analysis and visualization obsolete, resulting in disruptive changes to the scientific workflow and rendering traditional checkpoint/restart methods infeasible. Exascale system concurrency is expected to grow by five or six orders of magnitude, yet system memory and I/O bandwidth/persistent capacity are only expected to grow by one and two orders of magnitude, respectively. The reduced memory footprint per FLOP further complicates these problems, as does the move to a hierarchical memory structure. Scientific workflow currently depends on exporting simulation data off the supercomputer to persistent storage for post hoc analysis.

On Exascale systems, the power cost of data movement and the worsening I/O bottleneck will make it necessary for most simulation data to be analyzed in situ, or while the simulation is running. Furthermore, to meet power consumption and data bandwidth constraints, it will be necessary to sharply reduce the volume of data moved on the machine and especially the data that are exported to persistent storage. The combination of sharp data reduction and new analysis approaches heighten the importance of capturing data provenance (i.e., the record of what has been done to data) to support validation of results and post hoc data analysis and visualization. In this BOF, we will provide an overview of the efforts in ECP for supporting efficient and portable data management, analysis, and visualization on Exascale platforms.

The data management activities in ECP address the severe I/O bottleneck and challenges of data movement by providing and improving storage system software; workflow support including provenance capture; and methods of data collection, reduction, organization and discovery. We will report on the efforts and status of I/O libraries used by applications, HDF5, ADIOS, MPI-IO, and PnetCDF; the checkpointing library VeloC; the burst buffer file system UnifyFS; and data management support software developed as part of the DataLib project. All of these approaches are being designed and implemented in collaboration as part of ECP to address the needs of users of Exascale platforms.

Data analytics and visualization (DAV) are capabilities that enable scientific knowledge discovery. The ECP DAV portfolio is building an Exascale-capable ecosystem to support scientific discovery while addressing the challenges of analyzing, reducing, and visualizing data. In this section of the BOF, we will overview the ECP in situ infrastructures, Ascent and ParaView/Catalyst. The ALPINE project is developing a suite of in situ algorithms to enable automatic feature detection, selection and data reduction while the data is still resident in memory. These capabilities are built on top of VTK-m, the many-core visualization library used for cross-platform portability across the ECP DAV projects. We will also overview Interactive post hoc approaches that work on data extracts produced in situ to support post hoc scientific workflows through visualization tools such as ParaView, VisIt and Cinema. We will report on the status of the ZFP and SZ compression capabilities under development as part of ECP.

Attendees of our BOF will come away with an understanding of different approaches they can use to tackle their data management needs, avoid I/O bottlenecks and develop scientific workflows supported by both in situ and post hoc data analytics and visualization tools. Many of the software technologies overviewed will have in-depth BOFs later in the week.

Speakers:

Jim Ahrens (LANL)
Members of the ECP DAV portfolio

March 30, 1:00 pm - Observing GPU performance using the TAU Performance System

Abstract:

The TAU Performance System(R) [http://tau.uoregon.edu] is a powerful and highly versatile profiling and tracing toolkit for performance analysis of parallel programs at extreme scales. It supports performance evaluation of applications running on NVIDIA, AMD, and Intel GPUs supporting node-level programming models such as CUDA, OpenCL, OpenMP, SYCL (from OneAPI/DPC++), Kokkos, and Raja without any modifications to the application binary. This PIP meeting will highlight recent updates to TAU to support GPUs and discuss how TAU can be integrated in a performance engineering workflow to identify performance bottlenecks. TAU is integrated in the Extreme-scale Scientific Software Stack (E4S) and is available for download using the Spack package manager.

Speaker:

Sameer Shende (University of Oregon)

March 30, 1:00 pm - Open MPI State of the Union

Abstract:

Open MPI continues to drive the start of the art in HPC. This year, we’ve added new features, fixed bugs, improved performance, and collaborated with many across the HPC community. We’ll discuss what Open MPI has accomplished over the past year, and present a roadmap for the next year.

In this BOF, we’ll present where Open MPI currently is and where it is going. Here are some of the highlights of what will be covered during the discussions:

The status of the v4.1.x release series
Overview of the upcoming v5.0.x release series, including PRRTE
Plans for MPI-4.0 support
Improvements on collective operation performance, one-sided operations
Multithreading and asynchronous progress support
Support for Exascale machines

Be part of the discussion: submit your questions ahead of time and come learn where Open MPI is going, and how you can (and should!) join our efforts.

Speakers:

George Bosilca (The University of Tennessee)
Jeffrey M. Squyres (Cisco)
Joshua Ladd (NVIDIA)
Howard Pritchard (LANL)

March 30, 3:00 pm - HDF5 Community BOF

Slides from the Event:

ECP-HDF5-Comm-BOF-All-Slides

Abstract:

In this Community BOF, the HDF5 team will briefly present the latest features in HDF5 that improve I/O performance on large-scale systems. The BOF also includes presentations from ECP AD teams, such as EQSIM, AMReX, etc. on how they use HDF5 and discuss any performance issues. The floor will then be open for all HDF5 users (ECP Application Development and Software Technology teams) to discuss their usage, performance issues, new feature requests, etc.

Speakers:

Suren Byna (LBL)
Quincey Koziol (LBNL)
Houjun Tang (LBNL)
Scot Breitenfels (The HDF Group)
Elena Pourmal (The HDF Group)

March 31, 11:00 am - Updates and Roadmap for the PMIx Community

Slides from the Event:

https://github.com/pmix/pmix-standard/wiki/Event-2021-03-31

Abstract:

In this BOF, we will discuss the status of the PMIx Interface, the efforts of the PMIx Standard Body, the OpenPMIx implementation of PMIx and related projects, and to invite any interested attendees to participate in the development of PMIx.

The PMIx Standard is a formalization of the APIs previously developed by the OpenPMIx development community and aims to provide an implementation-independent specification to facilitate interaction between applications, tools, middleware, and system runtime environments. The interface focuses on supporting HPC application launch and wire-up; supporting tools that wish to inspect, steer, and/or debug parallel applications; providing interfaces to manage dynamic workload applications, to support fault tolerance and cross-library coordination; and facilitate communication across container boundaries. The BoF will cover in-progress proposed additions to the PMIx Standard, including support for I/O and storage and dynamic workflows; additions that have been formally adopted, including support for tools and debuggers; and updates and new features of the OpenPMIx implementation.

We will present the v4.0 PMIx Standard release and associated use cases, working groups will discuss their achievements and plans for the next year, then we will discuss the priorities and roadmap for the next year. Planned topics include:

New and proposed features for PMIx
Improvements and progress in OpenPMIx and related projects (e.g., PRRTE)
Progress in the PMIx Standard
Solicitation for proposed new features, use case scenarios, and working groups

The entire HPC community (and beyond!) can benefit from the standardized APIs included in the PMIx Standard and the OpenPMIx implementation. In this BoF, we hope to encourage broad community participation in this effort, including folks working on parallel libraries (e.g., MPI, UPC, OpenSHMEM), runtime environments (e.g., SLURM, LSF, Torque), container runtimes and orchestration environments (e.g., Singularity, CharlieCloud, Docker, Kubernetes), tools (e.g., TotalView, DDT, STAT), middleware (e.g., I/O support libraries, workflow management frameworks), and the broader research community. The PMIx Community BoF is an opportunity for the ECP community to influence the direction of the PMIx roadmap by providing valuable feedback on the current state and future direction of the PMIx Standard.

Speakers:

Kathryn Mohror (LLNL)
Thomas Naughton (ORNL)
Josh Hursey (IBM)
Members of the PMIx community

March 31, 11:00 am - Enabling Developer Productivity with Software Process Improvement

Abstract:

This BoF addresses the challenge of improving productivity in the development and use of scientific software. The organizers will provide a brief overview of the Productivity and Sustainability Improvement Planning (PSIP) process, a lightweight workflow for software process improvement developed by the IDEAS-ECP project to support ECP teams. PSIP captures the tacit, more subjective aspects of team collaboration, workflow planning, and progress tracking. In the potential absence of planning and development processes, and as scientific software teams scale to larger, more diverse, aggregate teams, unforeseen disruptions or inefficiencies can often impede productivity and innovation. PSIP is designed to bootstrap aggregate team capabilities into best practices, introduce the application of appropriate resources, and encourage teams to adopt a culture of process improvement. After a 10 minute introductory talk, we will open the floor to general discussion on how teams make progress in sustainable ways. Participants will learn from each other about how to approach teams, collaborators, and end users to precipitate process improvement.

Speakers:

Elaine Raybourn (SNL)
Reed Milewicz ( SNL)
Greg Watson (ORNL)
Elsa Gonsiorowski (LLNL)
David Rogers (ORNL)
Ben Sims (LANL)
Jim Willenbring (SNL)

March 31, 1:00 pm - MPICH for Exascale: Supporting MPI-4 and ECP

Slides from the Event:

Abstract:

This BoF session will provide a forum for users of MPICH as well as developers discuss experiences and issues in using and porting MPICH. The BoF will provide information on MPICH’s status of MPI-4 and GPU support. Future plans for MPICH will also be discussed. MPICH developers will also be present for an open forum discussion.

Speakers:

Yanfei Guo (ANL)
Ken Raffenetti (ANL)
Hui Zhou (ANL)

March 31, 1:00 pm - Cultivating Software Sustainability, Productivity and Quality through BSSw.io

Abstract:

The Better Scientific Software site (https://bssw.io) provides a central hub to share information on practices, techniques, experiences, and tools to improve developer productivity and software sustainability, as key aspects of increasing overall scientific productivity. BSSw.io features ever-expanding resources on scientific software planning, development, performance, reliability, collaboration, and skills. A GitHub backend enables content development using a collaborative, open workflow; content can also be contributed with an easy-to-use Google form.

In this BOF session, the organizers and the community will share experiences on using and contributing to BSSw.io. An objective is open discussion to understand pressing needs, challenges, and potential roadblocks for software sustainability and productivity, in order to determine priorities for future directions in site content, usage, and contributions. The long-term vision for BSSw is to serve as an international community-driven and community-managed resource, with content and editorial processes provided by volunteers. We welcome everyone to participate.

Speakers:

Rinku Gupta (ANL)
David Bernholdt (ORNL)
Roscoe Bartlett (Sandia)
Jean Shuler (LLNL)
Paul Wolfenbarger (Sandia)

March 31, 3:00 pm - ECP Math Libraries Capabilities and Applications Engagement

Slides from the Event:

ECP-Math-Libraries-BOF-03-31-2021

Abstract:

ECP math libraries encapsulate a diverse range of cutting-edge scalable algorithms to support the needs of ECP applications. Partnerships between The Application Development (AD) teams and math library developers potentially offer the biggest payoffs for ECP, where AD teams motivate new functionalities in math libraries and also can customize library capabilities as needed for their projects. We would like to take this opportunity to re-examine (1) what capabilities ECP math libraries provide, (2) what are our strategies to make the ECP math libraries architecture-aware and application-aware, and (3) how to pursue application-specific customization, so that together we can more effectively work toward next-generation science goals. In the first half of the BOF, the seven focus areas leads will give overviews of their respective math library capabilities. In the second half, we will conduct breakout sessions for in-depth discussion of each topical area. The application teams will have an opportunity for conversations with the math library developers to explain their needs, particularly focus on circumstances of applications teams, with goals of advancing understanding on both ends and setting the stage for deeper conversations in the future.

Speakers:

Sherry Li (LBNL)
Lois Curfman McInnes (ANL)

April 1, 11:00 am - The LLVM Compiler Infrastructure --- Clang, Flang, OpenMP, and More

Abstract:

The LLVM compiler framework is a cornerstone of the ECP ST efforts. The open source compilers and tools developed as part of the LLVM framework are valued and used by various application and software technology teams. At the same time, all HPC vendors utilize LLVM technology in their portfolio which makes it an all the more critical part of the DOE software ecosystem.

Throughout the ECP there are numerous interactions with the LLVM framework. Among other things, we employ existing LLVM technology such as the sanitizers and libc++ in our applications, we prototype new HPC-focused functionality for standards like C++ or OpenMP, we build and incorporate production quality enhancements to the open source code base to use it via Clang and Flang and the vendor toolchains. All in all, LLVM plays a major role when it comes to software technology in ECP, even though it is often rebranded.

In this BOF we will provide participants a closer look into the interactions of ECP with the LLVM framework. Our speakers will give brief presentations followed by Q&A. We intend to have breaks between the talks to foster discussions. The goal is to highlight the LLVM related efforts towards engaging with AD and ST teams in order to build synergies, leverage expertise and tools, and foster conversations that guide future efforts.

Our speakers will cover a wide range of topics, including but not limited to: Flang/OpenMP/OpenACC/… implementation status, useful debug and development tools, DSLs, auto-tuning, and performance estimation.

Speaker:

Johannes Doerfert (ANL)
Alexis Perry-Holby (LANL)
Valentin Clement (ORNL)
Joel Denny (ORNL)
George Stelle (LANL)
Jeff Vetter (ORNL)
Fady Ghanim (ORNL)
Joseph Huber (ORNL)
Michael Kruse (ANL)
Shilei Tian (Stonybrook)
Jan Hückelheim (ANL)
Gokcen Kesto (PNNL)
Pat McCormick (LANL)
Giorgis Georgakoudis (LLNL)

April 1, 1:00 pm - How to Measure and Analyze the Performance of GPU-accelerated Code

Presentation(s)

PAPI_BOF_Presentation-pdf

Abstract:

This BoF will present three tools – TAU, HPCToolkit, and PAPI. These performance evaluation tools support GPUs from NVIDIA, AMD, and Intel. The BoF will present details of interfaces used to collect performance data, instrumentation mechanisms, the HPC runtimes supported, the general methodology to access hardware performance counter data and associate it with source constructs. It will describe how developers can use these tools effectively to tune their codes on all three GPU platforms. The BoF will solicit input from the community.

Speakers:

Sameer Shende (University of Oregon)
John Mellor-Crummey (Rice University)
Tony Castaldo (The University of Tennessee)

April 1, 3:00 pm - Tools for Data-driven Analysis and Improvement of HPC Scientific Software Development

Abstract:

The BOF considers the role of tools in supporting the productive development of high-performance scientific software. The organizers will introduce current approaches to using software development related data within ECP, including, but not limited to code, git, emails, issues, test results, and dependencies, to improve understanding of development processes and to help enhance developer productivity. After short introductory talks by the organizers in the first 15 minutes, we will open the floor to a general discussion of tool-related topics, with a focus on the potential positive impact on productivity and code quality. The organizers will synthesize a report summarizing the community input on what the most pressing automation needs are, as well as any novel technical approaches that warrant further investigation.

Speakers:

Boyana Norris (University of Oregon)
Reed Milewicz (SNL)
Greg Watson (ORNL)
Keita Teranishi (SNL)