On Mon, Mar 27, 2017 at 3:10 PM, Dan Holmes <[email protected]> wrote:
Hi Jeff,
I wrote some notes about our discussion. https://github.com/mpiwg-p2p/p2p-issues/wiki/notes-2017-03-27
Indeed, as we discussed in San Jose last year, I don't think we need to make all 36 A-F pairs. Fsend doesn't block on any thing, so it is effectively Ifsend(req=REQUEST_NULL). Synchronous freeing send (Fssend or Sfsend?) is valid and arguably useful for all of the reasons that Issend is. Ready freeing send (same naming quandry) is valid but since I've heard lots of criticism of ready send before, I won't try to defend it in this context. Arecv and Iarecv both make sense. And they take "void**" (identical to void* in C, of course) just like MPI_Alloc_mem (and not like MPI_Recv).
In short, the memory must be returned to the original owner for de-allocation.
MPI owns the memory. No stack or user heap allocators allowed. Nothing else is the slightest bit practical.
We mentioned an attach/detach method for using user memory.
There is no value in this. MPI can't turn arbitrary user memory into shared in general so defeats the primary purpose.
We also discussed using Fsend/Arecv to/from MPI_COMM_NULL to transfer ownership without needing another message.
This makes no sense. Please elaborate.
This needs careful thought - hence the suggestion for sequence diagrams.
I drew most of them at some point. Easy to do again. Jeff
Cheers, Dan.
On 27 Mar 2017, at 22:42, Jeff Hammond <[email protected]> wrote:
Actually, it has to be required in the literal sense.
Fsend frees memory. Has to know which allocator used. MPI_Alloc_mem is the only one MPI knows about.
Same in reverse for Arecv and MPI_Free_mem.
How else can you do it? We could register callbacks for memory allocator but that will almost certainly prevent every useful optimization.
Jeff
Sent from my iPhone
On Mar 27, 2017, at 8:57 AM, Dan Holmes <[email protected]> wrote:
“required” is possibly too strong. “advisable” would be closer to my expectation. “needed in order for MPI to enable all possible optimisations” is more wordy/precise.
Cheers, Dan.
On 27 Mar 2017, at 16:55, Jeff Hammond <[email protected]> wrote:
No. MPI_Alloc_mem and MPI_Free_mem were going to be required. At least that was my plan.
Sent from my iPhone
On Mar 27, 2017, at 8:51 AM, Jim Dinan <[email protected]> wrote:
Did we previously look at using MPI_Buffer_attach as a way to support allocate-and-recv?
On Mon, Mar 27, 2017 at 11:38 AM, Jeff Hammond <[email protected]> wrote:
I'm on vacation but I endorse other people doing stuff with Fsend-Arecv, since the slacker who owned it for the past two years hasn't made any progress. I think I copied all the relevant content to GitHub already.
Jeff
Sent from my iPhone
On Mar 27, 2017, at 8:16 AM, Dan Holmes <[email protected]> wrote:
Hi Jim, et al,
I was hoping to move the WG on to talking about Freeing-Send and Allocating-Receive (Fsend & Arecv).
I’d like to re-boot and refresh caches on that with a goal of presenting something (probably informally) next face-to-face meeting in June.
I’ll start the call and see how many turn up.
Cheers, Dan.
On 27 Mar 2017, at 15:28, Jim Dinan <[email protected]> wrote:
Hi All,
The info query proposal seems to be converging. I don't think there's anything to discuss on this topic this week.
Current status is that Pavan will check with vendors to make sure they are ok with updating MPICH to report the user's info key value instead of the effective value being used by the implementation. Assuming this is ok, it sounds like the proposal that was read at the last meeting is ready to move forward into voting.
Are there any other topics for discussion? If not, I think we can cancel today's meeting.
Cheers, ~Jim. _______________________________________________ mpiwg-p2p mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpiwg-p2p
-- The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.
_______________________________________________ mpiwg-p2p mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpiwg-p2p
mpiwg-p2p mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpiwg-p2p
_______________________________________________ mpiwg-p2p mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpiwg-p2p
_______________________________________________ mpiwg-p2p mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpiwg-p2p
The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.
_______________________________________________ mpiwg-p2p mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpiwg-p2p
_______________________________________________ mpiwg-p2p mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpiwg-p2p
The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.
_______________________________________________ mpiwg-p2p mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpiwg-p2p
-- Jeff Hammond [email protected] http://jeffhammond.github.io/