Ray: control, ownership, and per-node object stores

Ray: control, ownership, and per-node object stores An architecture diagram generated by Archify. Driver · ObjectRef owner (example) · Architecture component Driver ObjectRef owner (example) GCS · cluster metadata · Head: control shown GCS cluster metadata Worker A · raylet + tasks / actors · Node A Worker A raylet + tasks / actors Object store A · local shared memory · Node A Object store A local shared memory Worker B · raylet + tasks / actors · Node B Worker B raylet + tasks / actors Object store B · local shared memory · Node B Object store B local shared memory control lookup submit + dependency node / actor control local read / store transfer when needed Head: control shown Node A Node B Legend Processes Memory stores Driver

Responsibility boundaries

  • • Separate GCS cluster metadata from ObjectRef owner metadata
  • • Head raylet, local store, and user work are omitted for clarity

Resources and memory

  • • A single task/actor must fit one node
  • • Distinguish same-node NumPy views from cross-node copies