Scientific computing teams face a high-stakes decision when designing storage infrastructure: which open-source parallel filesystem fits their workloads? With growing diversity — from large-scale simulations and climate modeling to genomics, AI-assisted research, and multi-disciplinary analysis — the wrong choice can mean performance bottlenecks, heavy operational overhead, or unnecessary cost.
This talk offers a grounded comparison of three widely used open-source parallel filesystems:

• Lustre — the long-established leader in traditional HPC, known for high sequential throughput and strong MPI-IO support, typically deployed for high-performance scratch.
• CephFS — the POSIX file interface of Ceph, the unified software-defined storage platform; CephFS shares the same RADOS cluster as Ceph's object (RGW) and block (RBD) interfaces, making it attractive where file, object, and block are needed from one system, particularly in cloud-native and Kubernetes (Rook) environments.
• BeeGFS — a lightweight parallel filesystem valued for simpler deployment and management in mid-to-large scientific clusters.

The session works through the key decision factors:

• Performance characteristics across scientific I/O patterns (sequential vs. mixed).
• Scalability approaches and metadata handling.
• Operational complexity — deployment, monitoring, and maintenance effort.
• Hardware flexibility and integration with modern tooling (Slurm, Rook, Kubernetes).
• Suitability for hybrid architectures — for example, sites pairing Lustre for high-performance scratch with Ceph for home directory, and long-term storage.

Rather than crowning a single winner, the talk gives attendees a clear decision framework — mapping workload type, scale, operational capacity, and hardware constraints to the file storage that best fits — drawn from production usage patterns in scientific computing.