spark split dataset to clusters