[mpiwg-hybridpm] SC'25 meeting solicitation to discuss FAM as a memory kind

[Ghosh, Sayan]

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-alloc10.pdf__;Lw!!G_uCfscf7eWS!dY-grOiqItDwceid4ydilTYj6sdJ7zu4RfqMrGrKzdLFZEEyANbO-z3E5PwUl0ZZY4P6cE_fmdSsZ3W0bpKiSbCpuXbZq9EDUwJ2BQ$ ) 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_uCfscf7eWS!dY-grOiqItDwceid4ydilTYj6sdJ7zu4RfqMrGrKzdLFZEEyANbO-z3E5PwUl0ZZY4P6cE_fmdSsZ3W0bpKiSbCpuXbZq9GZTK19yw$ 


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-grOiqItDwceid4ydilTYj6sdJ7zu4RfqMrGrKzdLFZEEyANbO-z3E5PwUl0ZZY4P6cE_fmdSsZ3W0bpKiSbCpuXbZq9H2Q5EWRQ$ 


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-hybridpm/attachments/20251008/d1918ddd/attachment.html>