During range shuffle, there are a huge number of index lookups, thereby taking at least the half time of the total shuffle time. This is because it is difficult to know which data is stored on which machine, so each machine needs to ask to all machines including itself.
Unfortunately, this problem is difficult to resolved using statistics. With an assumption of even data distribution, every machine is expected to have the data of the whole range.
So, we need to investigate an alternative solution to reduce the frequent index lookups.