JJL paper brainstorm

Programming Model & Execution Model

• Asynchronous execution of "tasks" mapped dynamically to nodes/ranks
• Built-by-design to be highly interoperable with MPI
  • Low-overhead and simple transition between MPI and vt, expected to occur every timestep, both can be active and work together
  • Constructs in vt have symmetrical use in MPI (example: vt groups and MPI groups)
  • Safe MPI collectives through consensus amidst an asynchronous, dynamic scheduler (novel?)
• Completion and ordering
  • PendingSend/DependentSendChain: task graph can expand beyond what is ready-to-run (RT system has lookahead)
  • Control and data completion are intrinsically bound together through epochs
  • Epochs form a graph that groups sets of tasks in the task graph
  • Multi-algorithm, lightweight, scalable termination detection of epochs
    • Epoch graph re-structuring optimizations; transitive dependency elimination? termination.cc:904
    • IntegralSet data structure to compactly store a set of compact ranges of integers
• Collective Design
  • Many interfaces in vt have collective interfaces to simplify use for MPI users (and reduce rooted overhead)
  • CollectionChainSet: drive collections of tasks collectively and asynchronously alongside MPI
  • Collective group creation voting algorithm (novel?) for efficiently building spanning trees across a subset of nodes group_info_collective.cc:258

LB

• Load models
• Python simulator framework (LBAF)
• Algorithm improvements to GossipLB

Applications

• EMPIRE as a use case
  • Performance results for some key problems
  • Transition between MPI and VT each timestep
• NimbleSM?

Engineering & Design

• C++14 adherence
• Wide and older compiler support
• Static template registration pattern for safe handler dispatch across memory spaces