In org.apache.hadoop.tools.DistCp#deleteNonexisting we serialize FileStatus objects when the path is all we need.
The performance problem comes from org.apache.hadoop.fs.RawLocalFileSystem.RawLocalFileStatus#write which tries to retrieve file permissions by issuing a "ls -ld <path>" which is painfully slow.
Changed that to just serialize Path and not FileStatus.
To delete the files we invoke the "hadoop" command line tool with option "-rmr <path>". Again, for each file.
Changed that to dstfs.delete(path, true)