Side Document defining MPI_T entities

[mahermanns]

Problem

Currently, MPI does not mandate any MPI_T entities to be implemented which makes it hard for portable tools to work across different MPI implementations. Furthermore, some entities only make sense in the context of other entities as in an "all or nothing" approach.

Proposal

To retain MPI_T's philosophy of not mandating specific entities to be present, a side document should define specific entities or groups of entities with specific semantics. Then if an MPI implementation claims to implement a specific entity that is listed, it will have to follow the specified semantics.

Changes to the Text

tbd.

Impact on Implementations

Implementors can choose whether to implement a specific set of MPI_T entities or not.

Impact on Users

Tool developers could rely on specific semantics when an entity is discovered to be supported.

References and Pull Requests

[mahermanns]

Meeting Tools WG 2025-01-20:

• Defining optional entity sets with specific coherent semantics
• Maybe start with specification of event sets of what used to be MPI Peruse (and still is available in Open-MPI?)
• Functionality of MPI Advisor and TAU/MVAPICH2?
• Distributed Lock Tracking
• Sieving though available implementations and comparing semantics of similar entities
• We need to identify/contact/engage with MPI implementors for these discussions

[mahermanns]

Meeting Tools WG 2025-02-17

• @cchambreau prepared a PR to integrate PERUSE via MPI_T in OpenMPI
• @hppritcha has a PR with a stub implementation of MPI_T in OpenMPI
• Agreed to read up on old/current PERUSE spec, as well as Chris' PR to incorporate it into OpenMPI, then discuss at the March meeting in-person on how to proceed

[wrwilliams]

Meeting Tools WG 2025-03-17

TODO on the tools wiki as a draft:

• Map Peruse events to MPI_T event specification
  • Message/Pt2Pt
  • RMA

TODO, general:

• Consider generalizations of Peruse's Pt2Pt event set for clever collectives
• Consider which Peruse events have application as PVARs
• Consider callback safety requirements of Peruse events--use OMPI levels as a starting point