Jeff, Just to add more info for the O(N) argument. On IB, I implemented window dynamic and I exchange registration information on-demand. So the information is only maintained for the processes you actually communicate with. Sreeram Potluri On Fri, Sep 23, 2011 at 10:19 AM, Jeff Hammond <[email protected]>wrote:

This is one case where I do not care about progress :-) I just want to avoid deadlock. I'd be happy to enqueue all window creation calls and have a single collective that would let me fire all of them at once. I get that this is not really MPI style, but no one has convinced me that there is a portable, scalable way to solve this problem.

I can solve the O(N) metadata problem on BG, it just requires more work. Almasi has already done this, it just isn't in MPI. The motivation for putting windows on small groups instead of world is that this is sufficient. Why should I have to allocate a bloody window on world just to get around the blocking collective nature of window creation and that deadlock-ability that results?

Jeff

On Fri, Sep 23, 2011 at 1:27 AM, Barrett, Brian W <[email protected]> wrote:

...

In the Portals 4 implementation of dynamic windows, if the comm is MPI_COMM_WORLD (or has the same group as MPI_COMM_WORLD), there is no O(N) metadata with dynamic when the hardware supports a rational memory registration strategy (and we have NIC designs where this is the case). There is a single list entry and memory descriptor on each process, which shares a pre-defined portal table index. Over other networks, there may be some scaling of resources, but there are networks where that is not the case.

There will be come scaling of resources with size of group as sub-communicators are created, due to the need to map ranks to endpoints; this is a problem with subgroups in general and not windows or communicators specifically.

I think adding non-blocking window create is a bad idea; there are many synchronization points in creating a window, and I'm not happy with the concept of sticking that in a thread to meet the progress rules.

Brian

-- Brian W. Barrett Scalable System Software Group Sandia National Laboratories ________________________________________ From: [email protected] [ [email protected]] on behalf of Jeff Hammond [ [email protected]] Sent: Thursday, September 22, 2011 4:08 PM To: MPI 3.0 Remote Memory Access working group Subject: Re: [Mpi3-rma] nonblocking MPI_Win_create etc.?

The reason to put windows on subgroups is to avoid the O(N) metadata in the window associated with registered memory. For example, on BGP a window has an O(N) allocation for DCMF memregions. In the code my friend develops, N=300000 on comm_world but N<200 on a subgroup. He is at the limit of available memory, which is what motivated the use case for subgroup windows in the first place.

I do not see how one can avoid O(N) metadata with MPI_Win_create_dynamic on comm_world in the general case, unless one completely abandons RDMA. How exactly does registered memory become visible when the user calls MPI_Win_attach?

Jeff

On Thu, Sep 22, 2011 at 4:58 PM, Rajeev Thakur <[email protected]> wrote:

...

In the new RMA, he could just call MPI_Win_create_dynamic once on comm_world and then locally attach memory to it using MPI_Win_attach. (And avoid using fence synchronization.)

Rajeev

On Sep 22, 2011, at 4:25 PM, Jeff Hammond wrote:

...

I work with someone who has a use case for nonblocking window creation because can get into a deadlock situation unless he does a lot of bookkeeping. He's creating windows on subgroups of world that can (will) overlap. In order to prevent deadlock, he will have to do a global collective and figure out how to order all of the window creation calls so that they do not deadlock, or in the case where that requires solving an NP-hard problem (it smells like the scheduling problem to me) or requires too much storage to be practical (he works at Juelich and regularly runs on 72 racks in VN mode), he will have to serialize window creation globally.

Nonblocking window creation and a waitall solves this problem.

Thoughts? I wonder if the semantics of nonblocking collectives - which do not have tags - are even sufficient in the general case.

Jeff

-- Jeff Hammond Argonne Leadership Computing Facility University of Chicago Computation Institute [email protected] / (630) 252-5381 http://www.linkedin.com/in/jeffhammond https://wiki.alcf.anl.gov/index.php/User:Jhammond _______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

-- Jeff Hammond Argonne Leadership Computing Facility University of Chicago Computation Institute [email protected] / (630) 252-5381 http://www.linkedin.com/in/jeffhammond https://wiki.alcf.anl.gov/index.php/User:Jhammond

_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

-- Jeff Hammond Argonne Leadership Computing Facility University of Chicago Computation Institute [email protected] / (630) 252-5381 http://www.linkedin.com/in/jeffhammond https://wiki.alcf.anl.gov/index.php/User:Jhammond

_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma