Storage decides whether expensive GPUs compute or wait. AI workloads mix millions of small random reads during data loading, bursts of very large sequential writes during checkpoints, metadata storms from tokenized shards, and increasingly the offload of inference KV cache to shared storage. Each of the leading parallel systems handles that mix differently, and each has tunables that matter and many that do not. Cluster-level storage tiering and sizing are introduced in inf-310; this course goes inside the file systems themselves.
You start with the I/O physics of AI workloads, then work through each platform as an administrator: IBM Storage Scale (NSDs, Erasure Code Edition and Storage Scale System, filesets, the policy engine, AFM and the data acceleration tier introduced in Storage Scale 6.0); Lustre (MDTs and OSTs, LNet, DNE, Progressive File Layouts, Data-on-MDT and the features of the 2.16 and 2.17 releases, including DDN EXAScaler packaging); WEKA NeuralMesh; and VAST Data's DASE architecture and its AI-oriented data services. Cross-platform modules cover GPUDirect Storage and RDMA clients, client-side tuning, benchmarking with fio, IOR, mdtest and the MLPerf Storage benchmark while defeating cache effects, and operations: tiering to object, quotas, snapshots, replication, upgrades and failure handling. Facts are stated as of October 2026.