Optimizations for Neighborhood Collectives

[pavanbalaji]

The current implementation of neighborhood collectives in MPICH is a very simple sequential isend/irecv where each process sends the data it needs to send to its neighbors. More sophisticated algorithms are possible and should be implemented.

[jain-surabhi-23]

There is a series of PRs - #3371, #3372, #3375, #3376, #3377, #3378 for neighborhood collectives. Does it apply here? The PRs implement some neighborhood collectives using transport (TSP) API.

[pavanbalaji]

@jain-surabhi-23: Is there a study that shows when these neighborhood collective algorithms do better or are these being blindly added?

[hzhou]

These prs are all linear algorithms rewrote into TSP based schedule.

[pavanbalaji]

So, just porting of the current code to TSP? That's not really an optimization and is unrelated to this issue.

[raffenet]

A basic optimization we have discussed could be implemented as follows.

1. For a given communication graph, internally create an array of subcommunicators for use in neighborhood algorithms. Each subcomm comprises a process and its incoming neighbors.
2. Implement new neighborhood collective algorithms using these internal subcomms.

We believe the benefit of these new algorithms should be evident with stencil codes where the number of neighbors is high. For now, ignore the expense of creating and storing the additional subcomms.

[Bayyapu]

@raffenet I am trying to note the FY19Q2 progress. Could you please let me know the update or point me to the correct PR that i need to see for the progress of this issue?

[raffenet]

There is no PR at this point. We should sync next week. I hope to put something together on Tues-Weds.