6.5. Server-Side Collective Tracking

PMIx collective operations (PMIx_Fence, PMIx_Connect, PMIx_Disconnect, and group construct/destruct) are assembled in two phases: the server first gathers the contributions of all local participants, then hands the assembled operation up to its host environment for global completion across nodes. This section describes how the server tracks local participation, why the historical counter-based accounting is unsafe when a participant is lost, and the identity-based tracking model that replaces it.

• 6.5.1. Specification: Tracking Local Collective Participation
  • 6.5.1.1. Scope
  • 6.5.1.2. Background: how a collective is assembled
  • 6.5.1.3. The counter-based completion test and its vulnerability
  • 6.5.1.4. Required model: track participation by identity
  • 6.5.1.5. Lost-connection accounting rules
  • 6.5.1.6. Required architecture: a single tracking service
  • 6.5.1.7. Additional required behaviors
  • 6.5.1.8. Invariants
  • 6.5.1.9. Relationship to the current implementation
• 6.5.2. Implementation Plan: A Single Collective-Tracking Service
  • 6.5.2.1. Objectives
  • 6.5.2.2. Non-goals
  • 6.5.2.3. Current state (starting point)
  • 6.5.2.4. Design
  • 6.5.2.5. Migration sequence
  • 6.5.2.6. Testing
  • 6.5.2.7. Risks and constraints
  • 6.5.2.8. Open decisions to resolve before coding