guest ::
| Home > Publications database > Portable, Streaming HPC Containers for the EBRAINS Software Distribution |
| Home > Publications database > Portable, Streaming HPC Containers for the EBRAINS Software Distribution |
| Poster (After Call) | FZJ-2026-05739 |
; ; ; ;
2026
Abstract: <p> Reproducibility in computational neuroscience requires more than just simulation code; it demands a reliable, end-to-end software stack. Modern neuroscience workflows often involve complex dependency chains—spanning MPI implementations, numerical libraries, and Python runtimes—that are difficult to deploy across heterogeneous HPC infrastructures. Differences in OS versions, site-specific software stacks, and hardware optimizations frequently break portability.</p><p> The EBRAINS Software Distribution (ESD) addresses these challenges by leveraging Spack to track and build reproducible environments. However, maintaining parity across cloud systems, local workstations, and HPC facilities remains a hurdle, particularly when navigating host-specific dependencies like MPI, parallel filesystems, and GPU acceleration.</p><p> To solve this, we introduce an Apptainer-based container strategy that enables zero-install ESD deployment for end users across diverse HPC environments. We support three primary deployment modes:</p><ul> <li> <strong>Fully self-contained images:</strong> For maximum portability in isolated environments. </li> <li> <strong>Slim containers via CernVM-FS (CVMFS):</strong> Using bind mounts to stream software on demand, significantly reducing image size and distribution overhead. </li> <li> <strong>Embedded CVMFS:</strong> Providing the benefits of streamed software even in facilities where host-side CVMFS is not available. </li></ul><p> While streaming optimizes storage and deployment, cold-cache access can introduce startup latency; we analyze this trade-off for short-lived and bursty workloads.</p><p> To ensure near-native performance, our containers are tightly integrated with the HPC runtime. We bundle OpenMPI, UCX, <code>libibverbs</code>, and PMIx clients, ensuring seamless interoperability with host-side SLURM schedulers and high-speed interconnects. For GPU workloads, our containers bridge host NVIDIA drivers with compatible user-space CUDA libraries, maintaining performance parity without requiring redundant driver installation.</p><p> Beyond software portability, hardware heterogeneity introduces additional challenges, particularly across CPU microarchitectures with differing AVX2 and AVX-512 support on Intel and AMD nodes. To address this, we use QEMU user-mode emulation during the Apptainer build process to produce optimized, architecture-specific binaries while preserving a unified container interface.</p><p> This work will be a core aspect of the ESD’s move toward streaming containers via CernVM-FS for on-demand deployment across HPC infrastructures.</p>
|
The record appears in these collections: |
