8.7. Scalable Collectives Plans

Plans and measurements for the DVM-wide collectives — restructuring grpcomm around operations rather than components, and the record of what the lateral-movement experiment built, measured, and why it was withdrawn.

• 8.7.1. Restructuring the DVM collectives around operations
  • 8.7.1.1. Context
  • 8.7.1.2. Status
  • 8.7.1.3. The operations
  • 8.7.1.4. Cost model
  • 8.7.1.5. Design
  • 8.7.1.6. What the numbers turned out to be
  • 8.7.1.7. What a client actually asks about a remote peer
  • 8.7.1.8. What Slurm’s PMI2 does, and what it tells us
  • 8.7.1.9. Reassessing the allgather, after the reset
  • 8.7.1.10. Verification
  • 8.7.1.11. Open questions