October 2025 - mpiwg-hybridpm - mailman3.cels.anl.gov

Meeting this week
by Joseph Schuchart 28 Oct '25

28 Oct '25

1 0

Re: [mpiwg-hybridpm] Call for topics
by Rohit Zambre 14 Oct '25

14 Oct '25

When I had looked at endpoints, I used this ticket from Jim as a reference for the Endpoints’ proposal — https://urldefense.us/v3/__https://github.com/mpi-forum/mpi-issues/issues/5… I think the endpoints solution seamlessly addresses a variety of use cases in hybrid MPI+X programming models . Regards, Rohit From: mpiwg-hybridpm <mpiwg-hybridpm-bounces(a)lists.mpi-forum.org> on behalf of Joseph Schuchart via mpiwg-hybridpm <mpiwg-hybridpm(a)lists.mpi-forum.org> Date: Monday, October 13, 2025 at 6:58 AM To: Hybrid working group mailing list <mpiwg-hybridpm(a)lists.mpi-forum.org> Cc: Joseph Schuchart <joseph.schuchart(a)stonybrook.edu> Subject: Re: [mpiwg-hybridpm] Call for topics External email: Use caution opening links or attachments This Message Is From an External Sender This message came from outside your organization. Multi-device collectives are something we need to look into. Endpoints seem a good starting point. Was there ever a full proposal for it? A quick search only brought up Jim's 2014 paper [1] and some related papers. I put it on the list of topics on the wiki [2]. We'll go over it this Wednesday. Cheers Joseph [1] https://urldefense.us/v3/__https://dl.acm.org/doi/abs/10.1177/1094342014548… [2] https://urldefense.us/v3/__https://github.com/mpiwg-hybrid/hybrid-issues/wi… On 10/11/25 04:55, Jeff Hammond wrote: > I am curious if people have any appetite to bring back endpoints, not > for the original purpose, but in order to support GPU ranks. This is > one of the features that MPI lacks relative to NCCL that limit its use > in some AI applications. Some AI users do not want to create >1 > process per node just to manage a GPU, because this makes their > pre/post-processing much harder. They have multithreaded code for > that already and don't want to rewrite it or play silly games with > affinity to try to make the application efficient while idling GPU > management processes. > > Jeff > > On Fri, Oct 10, 2025 at 5:34 PM Joseph Schuchart via mpiwg-hybridpm > <mpiwg-hybridpm(a)lists.mpi-forum.org> wrote: >> Hi all, With the September Forum meeting behind us we should pick us the meetings again. Please let me know if you have any topics you want to discuss. We plan to call for joint meeting with the RMA WG to discuss device-side operations but the >> ZjQcmQRYFpfptBannerStart >> This Message Is From an External Sender >> This message came from outside your organization. >> >> ZjQcmQRYFpfptBannerEnd >> >> Hi all, >> >> With the September Forum meeting behind us we should pick us the >> meetings again. Please let me know if you have any topics you want to >> discuss. >> >> We plan to call for joint meeting with the RMA WG to discuss device-side >> operations but the RMA WG first needs to go through the feedback from >> the Forum meeting. Given that people are busy with SC prep, it will >> probably slip into December. >> >> Other topics are of course welcome :) >> >> Thanks >> Joseph >> >> _______________________________________________ >> mpiwg-hybridpm mailing list >> mpiwg-hybridpm(a)lists.mpi-forum.org >> https://urldefense.us/v3/__https://protect.checkpoint.com/v2/r01/___https:/… > >

1 0

Meeting this week
by Joseph Schuchart 13 Oct '25

13 Oct '25

1 0

Re: [mpiwg-hybridpm] Call for topics
by Joseph Schuchart 13 Oct '25

13 Oct '25

1 0

Re: [mpiwg-hybridpm] Call for topics
by Jeff Hammond 11 Oct '25

11 Oct '25

1 0

Call for topics
by Joseph Schuchart 10 Oct '25

10 Oct '25

1 0

SC'25 meeting solicitation to discuss FAM as a memory kind
by Ghosh, Sayan 08 Oct '25

08 Oct '25

Dear folks, We are trying to organize a 1-hour F2F meeting during SC’25 to discuss the prospect of FAM (Fabric Attached Memory) memory kinds in MPI. Currently, the v1.0 side-document on memory allocation kinds (https://urldefense.us/v3/__https://www.mpi-forum.org/docs/*sidedocs/mem-all… ) considers GPU/device memory kind. We want to consider proposing “FAM” as a legitimate memory allocation kind and discuss the challenges. For e.g., MPI buffers can be resident on FAM, allowing load-store access instead of data movement over RDMA. Current technologies such as CXL 2.0 enabled memory pooling requires software managed coherency for multi-hosts, which means users must invoke a sequence of clflush/fence instructions or use locks. Availability poll is below; please respond to me directly so I have your email. https://urldefense.us/v3/__https://www.when2meet.com/?32833111-SEjB9__;!!G_… Thanks! -- Sayan Ghosh Computer Scientist Future Computing Technologies Group Pacific Northwest National Laboratory (509)-372-6323 • https://urldefense.us/v3/__https://sg0.github.io__;!!G_uCfscf7eWS!dY-grOiqI…

1 0