Performance

• Batched reads (get_batch(), the DistDataset default): 10–27× cheaper per row for small rows; 1.2–3.9× faster VAE epochs.

• Loader workers: one or two ThreadDataLoader workers usually hide batched reads; forked DataLoader workers need persistent_workers=True.

• Reuse destination buffers with register_recv(); fresh buffers are registered on every read (get_profile(name)["mr_miss"]).

• method=1 is fastest; batched method=0 comes close for small rows.

• Profiling: DDSTORE_PROFILE=1 + get_profile(name) splits read time into lock wait, registration, fi_read post and completion, and GPU sync (vae-ddp.py prints a summary). bench_get.py measures per-row cost vs row size, destination, batch size and threads.

Measurements: results.