Why LCS create one large temp file in one disk when add New Node or after rebuild

Cassandra 2.1.15

we use LCS  12*2T disks  ,   2~3T per node.

there is one problem after we add one new node or rebuild from anthoer DC.

At the new node.

the SSTables in each level will changes like this:

step 1:  [12341/4, 0, 0, 0, 0, 0, 0, 0, 0]

=>

step 2:  [18/4, 10, 17, 0, 0, 0, 0, 0, 0]

=>

step 3: [0, 10, 103/100, 1004/1000, 1622, 0, 0, 0, 0]

during the process, there will be one large compaction task in step 2.  and produce a large temporary file in only one disk.  then the large temporary  file will split many small files.

but sometimes, the large temporary  file is too large. the 2T disk is not enough for it.  then the compaction task failed.....