Re: [Mpi3-hybridpm] Hybrid WG meeting at the Forum
If it's implemented using BSD mmap then all nodes will need to give the same filename to map. My feeling is that if MPI can reliably generate the key/filename/whatever for us then it's unnecessary complexity to put this in the user interface. ~Jim. On 12/7/10 7:58 AM, Douglas Miller wrote:
In order to ensure all nodes get the same memory segment, they all have to use the same "key". At least if this is implemented on Unix SysV or POSIX shmem. Then there's the possibility of some nodes allocing on a sub-comm first, etc. I think its just avoiding a lot of potential complications and don't see what the downside is.
_______________________________________________ Douglas Miller BlueGene Messaging Development IBM Corp., Rochester, MN USA Bldg 030-2 A410 [email protected] Douglas Miller/Rochester/IBM
Inactive hide details for "Ron Brightwell" ---12/07/2010 12:14:38 AM---On 12/06/2010 05:59 PM, James Dinan wrote:"Ron Brightwell" ---12/07/2010 12:14:38 AM---On 12/06/2010 05:59 PM, James Dinan wrote:
*"Ron Brightwell" <[email protected]>* Sent by: [email protected]
12/07/2010 12:11 AM Please respond to [email protected]
To
"[email protected]" <[email protected]>
cc
Subject
Re: [Mpi3-hybridpm] Hybrid WG meeting at the Forum
On 12/06/2010 05:59 PM, James Dinan wrote:
On 12/6/10 5:49 PM, Pavan Balaji wrote:
Hi Doug,
On 12/06/2010 11:42 AM, Douglas Miller wrote:
Re: SHMEM ALLOC: I thought we were considering adding a "const char *key" param so that shared memory allocations better match OS (e.g. POSIX) shmem allocations and also to allow the call to be non-synchronizing.
As I recalled, this was still being discussed. Did we decide on adding the key? Ron?
I'll add it in as a discussion item.
I think that the key is primarily helpful as a bootstrap mechanism for processes to ensure they connect to the right shared segment. I think the intent was to have shmem alloc be collective so this should be extraneous.
Right. I think you lose several important optimizations for shared memory if the call is not collective. It doesn't have to be synchronous, but it does have to be collective.
-Ron
_______________________________________________ Mpi3-hybridpm mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-hybridpm
_______________________________________________ Mpi3-hybridpm mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-hybridpm
participants (1)
-
James Dinan