Problem 1746

The current strategy to prepare seeds for MPI has a logic flaw when used in conjunction with MT. Even if technically we do not observe issues it is theoretically possible to have all ranks producing exactly the same events (e.g. reuse of seeds).

The problem is the following:
MPI prepares the seeds for the ranks in a buffer so that:
We start from a master Seed, than the ranks seeds are prepared:
Rank#0 <-> Seed 0
Rank#1 <-> Seed 1
Rank#2 <-> Seed 2
and so on

However with MT each Rank is a MT run that has the master seeded from the MPI seed. From this master a per-event seed is generated in a similar way and there is no guarantee that the seeds are re-used. This situation could theoretically happen:
Rank#0 Event#0 <-> Seed 0
Rank#0 Event#1 <-> Seed 1
Rank#0 Event#2 <-> Seed 2
...
Rank#1 Event#0 <-> Seed 1 (The same as Rank#0,Event#1!!!
Rank#1 Event#1 <-> Seed 2 (The same as Rank#0,Event#2!!!

Practically this does not happen only because with MT we generate couples of seeds for the RNG engines.

The correct strategy would be to modify the strategy in MPI for which the per-rank seeds are staggered taking into account the number of events each rank will simulate. For example, if each rank will simulate 100 events:
Rank#0 <-> generate Seed 0,1,....,99 <-> Seed master thread with Seed 0
Rank#1 <-> Seed 100,101,...,199 <-> Seed master thread with Seed 100
...

I propose to make this change for 10.2
Andrea