Miek Gieben
TIL that having an #NFS server (on ZFS, but don't think that matters) and a 1000+ exports fails miserable... rcp.mountd goes crazy #linux
Jul 10 at 1:33pmWeb
Show threadkasperd2d ago

Which scenario are you dealing with?

  • One server many clients
  • One client many servers
  • One client with many mounts from the same server
Show threadMiek Gieben2d ago
@kasperd one (a few) servers with many clients
Show threadSandor Szücs2d ago

@miek @kasperd I can tell you the story of exposing logs to developers and QA at Zalando >12y ago.
We used nfs from all servers to mount logs to “loghost”. At some scale we had to restart the server once a week we moved all nfs to sshfs because user space is so much better to deal with than kernel space if you consider hanging I/O.
I also learned that systems with load >1000 can be very responsive. If you have 1000 processes hanging in an nfs call it doesn’t matter to the responsiveness of a system….

Before anyone says what a crazy idea: it worked well for long time. Simple and easy.

Show threadkasperd1d ago

I haven’t managed NFS on that scale myself. But that sounds like exactly the kind of scenario where NFS is supposed to shine.

I was under the impression that NFS was designed for the server to not need state for each client.