I am starting to think that I am obliged to use one core for every language and all content trees and do filtering on rootNodeId to search only in content related to current website.
There is a non-documented option in updatesearchindex script for top_node_id but I am not sure that it will work nice after the first index.
Anyway, 1 core for every language is good enough. If you share content in db, there is no reason to divide it in the index. Filtering on root node_id is easy to implement.