Image showing the logos for Spack and E4S and QR codes to their websites
The Spack package management system provides many capabilities needed for high-performance computing software. E4S is a curated collection of some of the latest open-source software libraries and tools for addressing scientific problems using AI and modeling and simulation, particularly focused on utilizing AI accelerators to advance scientific discovery. E4S relies upon Spack to manage its portfolio.

The Department of Defense (DoD) High-Performance Computing (HPC) Modernization Program recently conducted an independent study to evaluate the use of Spack and E4S (Extreme-Scale Scientific Software Stack) in delivering scientific libraries and tools to its user base. This study, conducted without direct assistance from the Spack or E4S teams, aimed to assess not only the quality and utility of the installed software but also the installation process itself. The findings highlight the potential of these tools to enhance the accessibility and performance of DoD’s HPC resources.

Phased approach to assessment

The study occurred in two distinct phases:

  • Phase 1 focused on using Spack to install 100 primary E4S software products from (commonly used by DoD users) across five HPC systems, leveraging four compiler systems and two implementations of the Message Passing Interface (MPI). Dependencies for these products were also included.
  • Phase 2 involved a deeper evaluation of 16 primary products from E4S that are widely used by the DoD HPC community.

The overall experience was overwhelmingly positive. Although there was an initial learning curve with Spack, the tool proved to be straightforward and efficient once mastered. Spack’s pre-configured recipes and parameter options from E4S further streamlined the installation process.

Key highlights and success stories

The study revealed several noteworthy achievements, showcasing the value of Spack and E4S in managing complex HPC environments.

Flux job scheduler: Spack facilitated the effortless installation of Flux, a modern job scheduler. Unlike traditional schedulers like PBS, Flux offers a more versatile design for managing workflows. Users can launch Flux directly from their accounts, enabling personalized node management within the broader system scheduler. Initial feedback on Flux’s functionality and ease of use was highly positive.

TAU performance analysis tool: The DoD found TAU, a widely used performance analysis tool, easy to install through Spack and E4S. The process supported various configurations, including setups for GPU environments. This flexibility simplifies the adoption of TAU across diverse computing platforms.

Kokkos performance portability system: Spack also enabled seamless installation of Kokkos, a performance portability system developed in part under the Department of Energy’s Exascale Computing Project. Kokkos has proven crucial for creating applications that can run efficiently across different parallel architectures. Although Kokkos is not yet widely used within the DoD, its role is expected to grow as the organization expands its use of GPUs from vendors like AMD and embraces future accelerators. The study confirmed that Kokkos worked flawlessly on existing CPU and GPU platforms after installation with Spack and E4S.

LAMMPS molecular dynamics software: Perhaps the most striking success was with LAMMPS, a classical molecular dynamics code used for materials modeling. Previously, DoD teams relied on their own installation methods. By adopting the Spack and E4S recipe, they achieved not only faster installations across multiple systems but also a twofold performance improvement in the resulting executable. The Spack recipe provided by the LAMMPS team included optimized configuration parameters tailored for HPC systems, demonstrating the tangible performance benefits of this approach.

Key takeaways

This independent study underscores the value of open-source ecosystems like Spack and E4S in the HPC landscape. Key benefits observed include:

  • Simplified software installation: Spack’s pre-configured recipes from E4S reduced installation complexity.
  • Performance gains: Optimized configurations delivered faster and more efficient executable code.
  • Future-proofing: Tools like Kokkos prepare the DoD for a broader range of parallel computing architectures.

Conclusion

The DoD HPC Modernization Program’s study showcases the transformative potential of Spack and E4S in streamlining software deployment and enhancing performance in high-stakes environments. As the DoD and other organizations embrace these tools, they simplify workflows and position themselves for future advancements in HPC technology. The experience reinforces the importance of fostering open-source ecosystems to meet the evolving demands of high-performance computing.

Software mentioned: Spack, E4S, TAU, Kokkos

CASS members supporting the software: PESO, RAPIDS, S4PST, STEP

Author: Michael A. Heroux
Published: January 10, 2025