8.2. Mapping by Device Plans
Design and implementation plans for placing processes against devices —
--map-by device=gpu and its relatives — rather than against a CPU object
that happens to sit near one. Source: Open MPI issue #14169.
- 8.2.1. Mapping and binding by device
- 8.2.1.1. Status
- 8.2.1.2. The request
- 8.2.1.3. What is in the tree today
- 8.2.1.4. Evidence: the topology the requester supplied
- 8.2.1.5. Design
- 8.2.1.6. Where the code goes
- 8.2.1.7. Deviations from the requester’s table
- 8.2.1.8. Testing
- 8.2.1.9. Decisions already taken
- 8.2.1.10. Decisions still to make
- 8.2.1.11. Out of scope
- 8.2.2. Device mapping: implementation plan
- 8.2.2.1. Two refinements to the design
- 8.2.2.2. Current state
- 8.2.2.3. Dependency graph
- 8.2.2.4. Phase A — remove the dead
distremnants - 8.2.2.5. Phase B — parameterize
byobjover its object enumerator - 8.2.2.6. Phase C — the PMIx device enumerator
- 8.2.2.7. Phase D — wrapper, directive, mapper, bind ceiling
- 8.2.2.8. Phase D2 — the user documentation
- 8.2.2.9. Phase E — the
interleavequalifier - 8.2.2.10. Phase F — report the assignment
- 8.2.2.11. Phase G — the requester’s topology, the harness, the swarm
- 8.2.2.12. Phase H — vendor identity and the visibility envars
- 8.2.2.13. Phases as landed
- 8.2.2.14. Verification checklist
- 8.2.2.15. Risks and open items
- 8.2.2.16. Task checklist