could you possibly group your data differently in the map phase so that each reduce phase processes less?