Choice has to be done by benchmarking following technics (that have still to be developped or improved):

• interface migration
• scotch weighted graph coloration
• metis weighted graph coloration
• PT-Scotch weighted graph re-partitionning (not sure if Pt-scotch provides this option)
• ParMetis weighted graph re-partitionning

Following criterion have to be checked (in order of priority for parallel algo):

• interface migration (triangles/ edges / and points along old interfaces should not remains on the new interface because it requires more iterations to converge toward the fully adapted mesh)
• load-balancing (for various cases of adaptation: uniform, non-uniform, massive insertion, collapses...): for now load-balancing tries to balance the number of input node graphs, we may try to implement smarter terchnics to evaluate the amount of work / time that we will pass in the Mmg kernel (from the input mesh and input metric provided to Mmg)
• amount of graph edges that are cut (that leads to MPI comms)