How many workers to parallel bulk indexing in cluster?

If N is the best number, I must send N parallel bulk to one node or split them between all nodes?