Quoting the other post, the design is motivated by
Another, albeit more philosophical, motivation, is avoid appearing as a universal solution but actually limited to a single pattern (master/worker), despite MPI is much more flexible and versatile. The approach is rather simple, as we simply distribute the task of a population of ODEs among the cluster, regardless of its size, and we don’t lock worker nodes to waiting/working cycle(without this kind of limitation of map_rect, along with issues like 939 and 1165, Torsten’s MPI scheme could be compatible with Stan’s).
I’d be happy to talk about it in next meeting if anyone still interested.