In a mpi plugin there is no sync in coral scheduling

[pogmat]

In coral we launch possibily MPI functions from Taskflow's thread pool.

• MPI function needs a unique order among all the ranks and that this order must be the same of every rank. Otherwise deadlocks and UBs
• With the thread pool the order is decided runtime by the scheduler and is not replicable.

Even if so far so good this can break at every moment.

The issue boils down to the common lore: you use multi-thread inside multi-node and not vice versa.

[pogmat]

Proposed solution: @luca-heltai

if an MPI env is detected by coral the thread pool consist of just 1 thread.
This implies deterministic runtime execution.
this implies deterministic graph resolution.

The MPI program will take care about spawning threads inside ranks.