Poor globbing performance in nfs-mounted directory with lots of subdirectories.

[kaysond]

Running fish 3.6.1 on SLES 12 (with clean home dir). For additional context, please see IlanCosman/tide#420

I have an nfs-mounted directory with ~4000 subdirectories. If I do something like set -l glob /path/a*/, where /path/ has lots of subdirs starting with a, it takes upwards of 2 min to complete. Doing an ls /path even at its worst, never takes more than ~10s. I'm not sure what the globbing is doing that is slowing it down.

The globbing does not appear to cache the results between commands, even though the filesystem itself does. If I do set -l glob /path/a*/; set l glob /path/ab*/ where the result is the same number of directories, the second glob runs instantly. When I run the same command a second time, though, it still takes 2 minutes, even though its immediately after.

[faho]

What does strace say? This is really gonna be one where we would need to reduce the syscalls.

You want something like strace -c fish -c 'set glob /path/a*/'

[faho]

Sidenote:

The globbing does not appear to cache the results between commands, even though the filesystem itself does

Yeah, as a rule we don't do caching like that - in most cases the filesystem cache will do better than we do and cache eviction is terrible.

I don't believe a cache is the correct answer here or most other places and don't want to add it.

[faho]

Okay, you're gonna love this one, we checked if a file existed that we already knew existed.

Since we're now using the same number of syscalls as "ls", and it's all stat-ish (newfstatat vs statx), I'm gonna call this one done for now.

Thanks for filing!

[kaysond]

Okay, you're gonna love this one, we checked if a file existed that we already knew existed.

Since we're now using the same number of syscalls as "ls", and it's all stat-ish (newfstatat vs statx), I'm gonna call this one done for now.

Thanks for filing!

Saw the same thing in strace! Thanks for the quick fix!