Refactor MPI communication when size is unknown

[kks32]

Describe the feature Use MPI_Probe to determine the size of the message, instead of passing the size first followed by the message. https://www.rookiehpc.com/mpi/docs/mpi_probe.php

Describe alternatives The current implementation passes the size of the object (particles) before passing the message. This requires passing 2 messages.

[bodhinandach]

@kks32 By doing MPI_Probe and MPI_Get_count, do you need to block operation? Particularly, if you are putting the MPI_Send in a loop or iteration. Whereas when you send the size, I think you can just MPI_Wait by using the array of status at the end of the loop. Will MPI_Probe result in less efficient communication?

[capellil]

1) Using MPI_Probe will be more efficient than sending the size in a different message because it will not require an extra message to be sent. (Behind the scenes, when a sender emits a message, it also joins an envelope along with it, containing information about the message sent such as its size. MPI_Probe allows the user to query that envelope to get the size of the message. The overhead of that action is negligible; it merely consists in checking a variable in a structure, by contrast to sending an additional message.) 2) Yes, MPI_Probe will block until the message mentioned is actually sent by the sender (so that the envelope is sent to the receiver). The "unknown-size problem" is long known in MPI and distributed-memory systems in general, this is why techniques such as MPI_Probe have been developed some time ago. If the fact that MPI_Probe blocks is a problem, there is a non-blocking version MPI_Iprobe available: see https://www.rookiehpc.com/mpi/docs/mpi_iprobe.php

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[kks32]

Still todo, keeping this open.