Ceph for Big Science
summary
CERN runs Ceph in production since 2013 with eight clusters: 6 PB Cinder/Glance, CephFS at ~1.5 PB, RADOSGW S3 and >5 PB of physics objects, and continues BlueStore migrations. CephFS multi‑MDS is stable and serves HPC via OpenStack Manila; IO500 tests place the setup around rank ~10 with metadata as the main bottleneck. Hyper‑converged tests use ~300 compute nodes; tooling needs include ceph client/bd top, backup and pool‑level conversion, and large‑scale robustness proven to 10,000+ OSDs.
AI-written summary · about