This standard library call: nftw (or ftw)
IBM Knowledge Center

supports the find command traversing directory file trees - i.e., searching and locating files.

Assuming you want to keep the command you already have (and I am not sue that Rudi's suggested test is valid because of file and directory caching ):

A limiting factor is known to be the number of sub-directories in the file tree, and possibly the number of available open file descriptors - a per process limit.
If you can parallelize your code using several processes it may improve performance. I'm not sure this will help much because it depends on the number of sub-directories being large to gain any benefit. The developers who write system code try to maximize throughput.

What I'm saying is: performance enhancement work is subjective and often a misplaced resource and a waste of programmer time.
Suppose your command runs in one minute in production. Then you work hard and get it down to 35 seconds. The user perception of "slow" will still be there, so you have to get it down to maybe 6 seconds to make users happy and see it as "faster". In this case getting an order of magnitude improvement may not be possible.

And in this case you would have to do something about directory caching messing up testing because (you check this yourself) once you open a directory the system caches it for speedier access. Use the time command and rerun the command to see what I mean:

Code:

time [my long command goes here]
#write down the result
time [my long command goes here]
# write down the result and compare the two resulting times