-
Notifications
You must be signed in to change notification settings - Fork 10
July 1 2024
Carl Pearson edited this page Jul 1, 2024
·
3 revisions
Attds: Hugo, Evan, Gabriel, Joseph, Carl, Matthew, Stephen, Vivek
Carl
- still a bit hung up on https://github.com/kokkos/kokkos-comm/pull/81 (I think?) with regards to
Kokkos::View
ofstruct
, and creating / releasing an MPI Derived Datatype- decided to table the init/finalize stuff last week
-
https://github.com/kokkos/kokkos-comm/pull/100
- (Carl) start with 1st piece as it's own PR - keep send/recv in the API, move MPI stuff to KC::Mpi
- 3rd piece:
plan
is an optimization point. Maybe very minimal API supported insideplan
. - (Joseph):
group
instead ofplan
? orsubmit
? - (Joseph): returning vector of requests is challenging because if we do optimizations the number of requests and what each one means may change. Maybe just stream ordered, drop requests entirely - infrastructure in Kokkos may not exist for this?
- (Cedric): Kokkos graphs are experimental and might change a lot. Let's break up the PR.
- https://github.com/kokkos/kokkos-comm/pull/89 needs a review (documentation)
- The name
- put your suggestion in slack or the github discussion, we will have a formal vote this week
- https://github.com/kokkos/kokkos-comm/discussions/36
- NCCL (Gabriel, Nicole, Evan)
- Gabriel
- I can try send/recv once the Transport PR is up
- NCCL missing collectives and how to implement it https://github.com/dssgabriel/sampik/blob/channels/docs/NCCL_OVERVIEW.md#example-implementation-of-collectives-using-p2p-calls
- Some NCCL stuff is quite similar but not exactly the same as MPI (MPI_IN_PLACE, MPI_ANY_SOURCE). MPI is not required to initialize NCCL, you just need a way to communicate unique IDs between ranks.
- CI for GPUs (Carl)
- no update, waiting on name
- Applications
- ArborX (Andrey)
- Modern C++ / MPI
- P2300 sender/receiver accepted in committee https://isocpp.org/files/papers/P2300R10.html
- Jospeh: interest in an API that fits P2300
- P2300 sender/receiver accepted in committee https://isocpp.org/files/papers/P2300R10.html
- Accelerator-Initiated Communication / Support
- Smart NICs
- Cedric: thinking about / mocking device-initiated communication with Gabriel
- Vivek: anything nice we can do with tools here? Let's not miss any opportunities.