MAIN SCHWARZ.

General.

Setup the push_one_by_one window to push directly from the solution vector.

MPI_Init_thread ?

BUGS:

onesided global tree checking has a bug.

OPTI:

See if it is possible to directly communicate from the local vectors instead of gathering to the sol vec using the local_to_global function.

GPU: Local_solution needs to be copied to the CPU send buffer and MPI_Put. Use Gather and Scatter operations for the accumulation in exachange_onesided.

GPU onesided:

Check the tolerance checking if the squared tolerance needs to be passed in, especially in one-sided.

CPU threading:

Use threading with pthreads to setup the 1 MPI rank per node and multiple node.

Setting affinities:

Setup a HWLOC interface similar to MGPU/Ginkgo to set GPU/CPU affinities.

EVENT BASED.

?