Beyond a Single Stack: How Increasing Hardware Diversity Shapes Bunya’s Software Build Strategy

Harris S1

1The University Of Queensland, St Lucia, Australia

Biography:

Shelly is a Research Facility and Operations Officer at the University of Queensland.

Coming from a non-technical background in retail management, Shelly joined the Research Computing Centre at UQ in 2023 while completing a Bachelor of Computer Science.

Her areas of interest include scripting and automation, storage infrastructure, and scientific software builds.

In her spare time, she can be found crocheting, reading, or running.

Abstract:

Modern high-performance computing (HPC) systems are evolving into increasingly heterogeneous platforms to support the growing range and scale of research workloads. While leading exascale systems are typically homogeneous in design, smaller systems commonly incorporate a diverse range of hardware.

As HPC systems have become more diverse, maximising application performance without compromising consistency and reproducibility, and doing so in an operationally sustainable way has become increasingly challenging. Many organisations rely on a “build once, deploy everywhere” approach, with small team sizes often limiting the ability to support specialised builds.

Aimed at research infrastructure specialists, this presentation will explore the evolution of The University of Queensland (UQ)’s strategy for building HPC software in response to this challenge. UQ’s HPC, Bunya, currently consists of 12 distinct node types, with more being added shortly. Leveraging the EasyBuild framework, parallel builds, and a semi-automated build pipeline, the small team at UQ’s Research Computing Centre (RCC) builds and maintains a dedicated software stack for each node type.

The current iteration of the build process has reduced the time to rebuild the entire set of software stacks, each containing over 1000 separate modules, from several months to approximately three weeks.

Continual refinement of this process has significantly lessened the operational overhead of maintaining dedicated software stacks of this scale, but there is room for improvement. We will end by discussing the future of Bunya’s software strategy, before attendees are invited to share existing strategies and consider if a similar approach could enable greater optimisation at their sites.

 

Categories

Website Sponsor

Website Sponsor