Recently, I had to remove a directory containing 2.75 million files (some php debug blunder) in a vserver sub-directory of a machine which already ran under heavy I/O load. None of

rm -r $dir
find $dir -exec rm {} \;
ls $dir | xargs rm
[download]

was an option, since each would hog I/O, and the delay for productive tasks was unacceptable. Buffer cache was not an issue, plenty of memory being always available, and each large chain of multiple indirect blocks could be held in memory, processing each return from getdents(2) as it was delivered. Not so with ls, find et al, since those were hogging memory too and invalidating parts of the file system buffer cache while reading all entries.

Using perl, a readdir() loop, select() and unlink() solved that. My point is that readdir() gives you finer control than shelling out ls.