Optimizing multiple queries on scientific datasets with partial replicas