Nonresident attributes are stored in intervals of clusters called runs. Each run is represented by its starting cluster and its length. The starting cluster of a run is coded as an offset to the starting cluster of the previous run. Length and starting cluster are variable size fields. The first byte of a run indicates the size of both. The size of the offset is stored in the high nibble, and the size of the length in the low nibble.

For compressed or sparse runs, the offset is 0, and the size of the offset is also 0. Each compression unit starts at a multiple of 16 clusters. If compression is possible, at the VCN of a unit will be one or more data runs followed by an empty run. If there are data runs for more than 16 clusters, the unit was not compressible. If there is no data run at all (only a large empty run), the unit consists of all zeroes. 

Example: 21 20 ED 5 22 48 7 48 22 21 28 C8 DB
First run: 20 clusters starting from 5ED (5ED to 60D)
2nd run: 748 clusters starting from 5ED+2248 (2835 to 2F7D)
3rd run: 28 clusters starting from 2835+DBC8 (3FD to 425)
Note that the offset is interpreted as signed value.