We have been discussing this topic in the working group, and would be glad to have your input. If you're looking for something you can use today, NCCL is a good option. There's also a library called libmp that uses GPUDirect Async for stream-based communication [1].

 ~Jim.

[1] https://github.com/gpudirect/libmp

On Tue, Feb 9, 2021 at 3:09 PM Choi, Jaemin <jchoi157@illinois.edu> wrote:

Hi Jim,

 

Do you happen to know if there has been progress on supporting CUDA streams in the MPI standard?

I’d be interested to hear about this or other frameworks (e.g. NCCL) that support asynchronous communication via streams.

 

Best,

 

Jaemin Choi

PhD Candidate in Computer Science

Research Assistant at the Parallel Programming Laboratory

University of Illinois Urbana-Champaign

 

From: mpiwg-hybridpm <mpiwg-hybridpm-bounces@lists.mpi-forum.org> On Behalf Of Jim Dinan via mpiwg-hybridpm
Sent: Tuesday, February 9, 2021 1:10 PM
To: Hybrid working group mailing list <mpiwg-hybridpm@lists.mpi-forum.org>
Cc: Jim Dinan <james.dinan@gmail.com>
Subject: [mpiwg-hybridpm] Call for Topics

 

Hi All,

 

The Hybrid & Accelerator WG will meet tomorrow. The agenda so far is empty. Please let me know if you have any topics that you would like to discuss.

 

Cheers,

 ~Jim.