How about using 20 compaction threads, as they were using in their test case?
Test 3.
Also, can we look into this: