Two ways to sort externally are Multi-Line Merging Sort and Bucket Sort, both with two passes. The Bucket Sort burdens the CPU less and is more efficient, while its usage is restricted heavily by the High-Bit scheme that distributes records into subfiles: the keys have to be integers; the sizes of subfiles may vary too much; the number of subfiles cannot be chosen freely. Based on statistical theory, this paper presents a sample-seperators scheme to broaden the ussage of bucket sort algorithm. A brief discussion on the convergance of sample-seperator estimation is given and the probability to avoid memory overflow is calculated. This scheme enables the bucket sort algorithm to be applied in the SheenkSort system to win the 2003 PennySort (the Indy category) competition.