#1443 Introduce two-steps parallelism
Comments from both defunct Redmine and the Gitlab placeholder have been copied / pasted here; the date of the posts are thus off from several months
The idea is:
On a first computer (preferably with loads of RAM) run the initialization phase which determines how to balance ressources between mpi threads. This run should be sequential (so maybe Metis will be used instead of Parmetis). Then in the classic run, use the already pre-processed data. The aim is to avoid each mpi thread to load the whole problem into memory before reducing it: it could load its own part of the model without taking everythin in the start of the problem.
This is a very important modification: we will be able to tackle much bigger models after this change (provided we get access to a machine with enough RAM - but fortunately we have a couple of them).
Note: this ticket should also help in time to implement a proper restart mode: most of the mechanisms will be the same.