Brainstorming on R

Consumers/Produces

• Definition: express a resource set
• R_local: subset of R per broker rank
  • IMP consumes this
• Producers + Consumers
  • Scheduler produces
  • Scheduler consumes
    • Resources incrementally handed back from the exec
  • Exec consumes
    • Used to determine which ranks job shells need to be executed
    • Divides R into R_locals
  • Job shell consumes R (and jobspec)
    • Also might use R_local
  • Users produce
  • Utilities (user-facing interfaces)
    • What resources are available/idle
    • What resources are allocated to a job (still)
  • Potential consumers/producers
    • Resource monitoring
    • Dyanmic scheduling
• Relation to HWLOC
  • core ids should be hwloc logical ids
  • independent but R can be built from hwloc
  • R is a superset (supports power, etc)
• Needs to be extensible enough to support both R and R_local requirements

Requirements

• Should be versioned
• Representation needs to be scalable
  • Should support a compact representation
    • Two-level compression (first representation, second compression alg)
• Should support rank & slot as an attribute
  • Could hurt scalability and compactness, need to look later at optimizing this

General

• Recursive nature
  • R should be passed from parent to child
• Should be validated against/reconciled with hwloc
• Resources
  • Arbitrary types and attributes
• Should be expressible enough to represent a graph
  • Should be a superset of HWLOC
  • Can be optimize for specific graph configurations (multiple trees)
  • Should support arbitrary types of edges between resources (and attributes on those edges)

Exec System

• Required resources
  • Node
  • Core
  • Memory
  • GPUs
• Needs to be able to resolve the resource set of each rank
  • In the general case, maybe the node name can be used
  • In the non-general case, the scheduler can annotate resources with rank information
• Should emit jobid as well for optimization on scheduler-side (add to github issue about exec system)

Job shell

Utilities

Users

Scheduler

• Required resources
  • resources required by exec
• Can support/allow arbitrary resources too

Common operations

• Intersection
  • between two Rs
• Filter
  • based on a key or attribute
  • variants:
    • just the resources that have the key
    • resources that have the key plus their children (even if the children don’t have the key)
• Serialize and deserialize
  • Sanitize invalid attributes like rank and slot when passing to child/receiving from parent
• Insert/Add
• Delete/Remove
• Traversal

Plan of attack

• First implement the emitter in resource matching service
  • Emit simple examples as JSON
  • Then work on a reader
• Producer and consumer owners can play with it and experiment
  • Enumerating and annotating slots and ranks