Migrating to Spack for Dependency Resolution, Management, and Reproducibility
Managing dependencies across multiple programming languages is a critical and common challenge in HPC software. The Air Vehicle project of the DoD's Computational Research and Engineering Acquisition Tools and Environments element of the High Performance Computing Modernization Program (HPCMP CREATE™-AV) develops and supports physics-based simulation tools, such as Kestrel and Helios, to decrease acquisition lifecycle costs by detecting design flaws and performance anomalies in government and industry developed aircraft. Our organization's in-house legacy package manager, which lacks dependency resolution mechanisms, has caused friction in updating software stacks, delaying upgrades and deployments. Our new GPU projects further highlighted the limitations of the existing system and led to an evaluation of alternatives.

Spack, a modern package management tool specifically designed for HPC systems and scientific software, was chosen for our migration plans. This presentation outlines our migration process, including planning, testing, and deployment for both CPU and GPU projects. Challenges include migrating a diverse set of over 80 build and runtime dependencies, training developers, and implementing best practices while expanding institutional knowledge of Spack.

This talk will provide practical strategies for planning migrations to Spack, such as documenting existing build processes, using Spack community resources, and addressing software license management challenges. We highlight long-term benefits, including improved build reproducibility, support for additional architectures, and plans to automate upgrades of dependencies within CI/CD workflows.
IMPACT
Accomplishment: Migrated over 80 build and runtime dependencies from legacy package manager to Spack for GPU-based simulations; Result: Reduced environment upgrade and deployment times by 50%. Improved upgrade agility, reproducibility, and maintainability. Accelerated delivery of high-fidelity air vehicle simulations, directly supporting DoD acquisition lifecycle processes and objectives.
PRESENTER
Robert, Trigg
robert.d.trigg5.ctr@mail.mil
301-278-8145
HPCMP CREATE(TM)
CATEGORY
GPU usage for HPC
SECONDARY CATEGORY
Software Infrastructure & DevOps for HPC
SYSTEM(S) USED
Nautilus, Raider, Jean