One semantic that I have used before is to allow the implementation to return an ordering tag (optional) after a request is submitted (say for a PUT). This ordering tag can then be used in subsequent operations to enforce ordering w.r.t. the previous request. For some networks that provide ordering on a per connection basis, the ordering tag can simply map to a particular connection. If the implementation does not support ordering for a specific (or any) request then it can return a special "no order" tag. Not saying this is a good thing, just have used it before. - Galen On Jan 27, 2008, at 8:13 PM, Pavan Balaji wrote:
Oh, and I forgot to add that the RMA operations themselves should probably be allowed to set a "fence" flag as well, which would be equivalent to RMA operation + 0-byte send with special tag set. That is, they would be flushed out immediately, but will not require any interference from the remote process.
-- Pavan
On 01/27/2008 07:08 PM, Pavan Balaji wrote:
Rich, Jarek,
The ordering should definitely not be enforced in all cases. As I mentioned in the brief description below, the application should be able to inform the MPI stack to flush out previous requests through a "special tag". Consider an example where the sender wants to place some data in the receiver buffers through RMA operations and inform it that the data is written through a regular send operation -- this can be achieved by tagging the last send with the special tag, instead of having the receiver wait for the window to be closed before being able to use the data received through RMA operations. This is similar to fence, but on a process-pair basis, rather than on a window basis.
-- Pavan
On 01/27/2008 06:50 PM, Nieplocha, Jarek wrote:
Good point Rich. Enforcing ordering for all the cases say on networks with adaptive routing can be expensive and unnecessary.
Jarek
________________________________
From: [email protected] on behalf of Richard Graham Sent: Sun 1/27/2008 4:44 PM To: MPI 3 Remote Memory Operations Subject: Re: [mpi3-rma] Plans ?
Can you elaborate on this ?
What is the intent ?
Do you want to have ordering be required, or optional ? Seems like there are times where you would like to have ordering guarantees, and others in which you want the network to blast things through as fast a possible, w/o concern for ordering.
Rich
On 1/27/08 12:26 AM, "Pavan Balaji" <[email protected]> wrote:
I was planning to write a proposal on ordering and completion of requests. Specifically, to treat RMA operations as ordered with respect to how application writers perceive it. That is, if process A sends an RMA message to process B, and then sends a regular message to process B (with a special TAG, for instance), the MPI stack on process B should deliver both the RMA message as well as the regular message to the application. That is, it should not wait for the remote window to close before doing so.
However, it'll be good to have a telecon to bounce off ideas and get initial comments before we go off and spend time writing more detailed proposals.
Is a telecon being planned? I don't seem to have received any email about this.
-- Pavan
On 01/26/2008 10:43 PM, Rajeev Thakur wrote:
It would be good to know if anyone is planning to write a proposal on any topic for MPI-3 RMA.
Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Richard Graham Sent: Wednesday, January 23, 2008 10:18 AM To: [email protected] Subject: [mpi3-rma] Plans ?
What are the plans for this working group ? At the meeting last week there seemed to be quite a bit of interest in this topic, and it seemed like there could be at least 2 groups working on this. Seems like, if this is the case, it would be better to try and coordinate early on within the working group, rather than try and rationalize two or more well developed proposals. Any thoughts here ?
Rich
_______________________________________________ mpi3-rma mailing list [email protected] http://lists.cs.uiuc.edu/mailman/listinfo/mpi3-rma
_______________________________________________ mpi3-rma mailing list [email protected] http://lists.cs.uiuc.edu/mailman/listinfo/mpi3-rma
-- Pavan Balaji http://www.mcs.anl.gov/~balaji _______________________________________________ mpi3-rma mailing list [email protected] http://lists.cs.uiuc.edu/mailman/listinfo/mpi3-rma
_______________________________________________ mpi3-rma mailing list [email protected] http://lists.cs.uiuc.edu/mailman/listinfo/mpi3-rma
-- Pavan Balaji http://www.mcs.anl.gov/~balaji _______________________________________________ mpi3-rma mailing list [email protected] http://lists.cs.uiuc.edu/mailman/listinfo/mpi3-rma