Storage system¶
HoreKa 2 provides a multi-level storage architecture optimized for both usability and high-throughput computing. The system combines shared parallel storage for user and project data with node-local scratch storage for active job execution.
Shared file systems¶
The primary shared storage is based on IBM Spectrum Scale (GPFS), providing globally accessible home and project directories. This storage layer is designed for long-term data access and user workspaces. It supports the normal workflow of storing code, results, and project files in stable directories.
Job-local storage¶
Each compute node also provides local NVMe SSD storage, exposed to the job as $TMPDIR. This temporary storage is intended for intermediate files created during execution and can significantly improve I/O performance for workflows that generate large temporary datasets.
For multi-node jobs, a BeeGFS on Demand (BeeOND) file system may also be requested, allowing temporary shared storage for a specific job run.
Storage model¶
The overall storage concept is designed to separate:
- long-term user and project data,
- high-throughput scratch space for active work,
- and temporary runtime storage for individual jobs.
This layered model helps balance performance, capacity, and data management for research workloads on the cluster.
For details on quotas and storage policies, see Storage.