Deprecate MPI_Graph_create

[hjelmn]

Problem

The scalability of MPI_Graph_create() is limited due to the global nature of its arguments. We should deprecate it.

Proposal

Deprecate the following:

int MPI_Graph_create (MPI_Comm comm_old, int nnodes, const int index[],
const int edges[], int reorder, MPI_Comm *comm_graph);

int MPI_Graphdims_get(MPI_Comm comm, int *nnodes, int *nedges);

int MPI_Graph_get(MPI_Comm comm, int maxindex, int maxedges, int index[],
int edges[]);

MPI_GRAPH

Changes to the Text

Add MPI_Graph_create() and associated functions to the deprecated interfaces.

Impact on Implementations

None.

Impact on Users

Update code to the MPI_Dist_graph_create() routine instead.

References

Insert any internal (other issues) or external (websites, papers, etc.) references here.

[hjelmn]

@dholmes-epcc-ed-ac-uk Agreed on this. Here is the proposal.

[dholmes-epcc-ed-ac-uk]

In addition, MPI_TOPO_TEST has a possible return value of MPI_GRAPH, which should be deprecated.

We should not forget to deprecate MPI_GRAPH_NEIGHBORS_COUNT and MPI_GRAPH_NEIGHBORS (pointing the reader to the "DIST" versions instead).

We should also consider the fate of MPI_GRAPH_MAP. It's parameters are identical to MPI_GRAPH_CREATE and the Advice shows an example usage in terms of its relation to MPI_GRAPH_CREATE. However, there is no "DIST" version for this function. My suggestion would be to define MPI_DIST_GRAPH_MAP (and MPI_DIST_GRAPH_MAP_ADJACENT?) and deprecate MPI_GRAPH_MAP at the same time.

The example implementation of MPI_GRAPH_CREATE (MPI-3.1 pp313-314) is wrong (if used as user code) because the communicator resulting from the MPI_COMM_SPLIT will not have a topology attached to it. However, it indicates the possibility of arguing that MPI_GRAPH_CREATE is not needed because it is syntactic sugar over MPI_COMM_SPLIT (in addition to the scalability argument).

There is text throughout Chapter 7 that refers to MPI_GRAPH and MPI_GRAPH_CREATE. For example, in section 7.6 (MPI-3.1 p314 lines 33-39) when introducing neighbourhood collective operations and discussing the sequence of neighbours. In that example the text already states "use DIST instead" but should be strengthened to "because MPI_GRAPH_CREATE is deprecated".

Should we create and include an example showing how to do the replacement of MPI_GRAPH_CREATE with MPI_DIST_GRAPH_CREATE_ADJACENT? That is, some boiler-plate code that extracts adjacency information from the arguments that would have been given to MPI_GRAPH_CREATE and uses that to form the arguments for MPI_DIST_GRAPH_CREATE_ADJACENT?

[hjelmn]

Yes, I agree. We should probably put together some text showing how to move to dist graph. I will update the issue with the other functions that need to be removed.

BTW, should this go on the main MPI forum issues as well?

[dholmes-epcc-ed-ac-uk]

BTW, should this go on the main MPI forum issues as well?

Yes, I think this is already a topic ready for wider-than-WG discussion/action. We could/should aim to have text ready to formally read at the next face-to-face meeting.

[hjelmn]

Sure. It shouldn't be difficult to get the text ready by June. Should I open the mpi-forum Issue?

[hjelmn]

I am working on the first cut of the text now.

[hjelmn]

Drafts are complete. What is a good way to share them? Should I create a branch on the mpi-standard repo in the mpiwg-coll organization?

[dholmes-epcc-ed-ac-uk]

@hjelmn Yes, please create a branch in the mpiwg-coll/mpi-standard repo. Do you have sufficient access permissions? If not, @tonyskjellum or @wesbland can authorise you.

[hjelmn]

No, I had myself added to the RMA WG. Forgot that I probably should be added here as well. @wesbland Please add me to mpiwg-coll.

[tonyskjellum]

Will do Anthony Skjellum, PhD 205-807-4968

…

On Mar 9, 2018, at 1:28 PM, Nathan Hjelm ***@***.***> wrote: No, I had myself added to the RMA WG. Forgot that I probably should be added here as well. @wesbland Please add me to mpiwg-coll. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[hjelmn]

Draft text has been uploaded. https://github.com/mpiwg-coll/mpi-standard/tree/deprecate_mpi_graph

[hjelmn]

The ticket number will have to be updated with the final mpi-forum/mpi-issues ticket number once that is up.

The question I have is; is it worth trying to deprecate MPI_Graph in MPI-3.2 (if there is one) or leave it for MPI-4.0? The benefit I see of deprecation in MPI-3.2 is we can potentially remove it in MPI-4.0 and be done with it.

[dholmes-epcc-ed-ac-uk]

I think usual practice is to deprecate whenever but only remove things in a major version. The most recent deprecation (https://github.com/mpi-forum/mpi-standard/pull/26) was merged after MPI-3.1 so will be included in MPI-Next, whatever that is. We should just aim for as soon as possible!

[hjelmn]

Ok, I will update the text to show as deprecated in MPI-3.2 and we can update if MPI-3.2 does not happen.