Looking at the SMP version of pardiso I may be able to roughly get the maximum memory from iparm- (https://software.intel.com/en-us/node/521691), but these are not used for the cluster version ?
Is there a way to get the estimated memory needed for the solve phase, per process or total, for example, if not possible before the factorization phase ?
That's correct. These feature will be added later into one of the next updates. We will keep you inform when this will be released. wrt estimation -- currently, you may try to use SMP version and it would very rough estimation needed memory for each MPI process.