I always have all these parallel things take a configurable number of worker go routines, then try different numbers and run btop or iostat to see how close I am to being hardware bound. Never had a machine where the limit was anywhere close to the number of open FDs allowed. I guess I haven't proven this but my idea is having the generic scheduler have to sort 60k FDs or 2M (I have a lot of dup files in my "backup everything, randomly and repeatedly, since 1998 file system) FDs is going to thrash stuff more than just having go's runtime and my own code.