Re: NFS Server Comments
On Mon, 1 Mar 1999, Tim Buller wrote:
> 1) The NFS server support is better in 2.2.x. Is it better "enough" to
> implement in a production system? Are people doing this?
There are some patches that you will need for kNFS to be "bulletproof"
particularly when working with Solaris. VAResearch is using it with a very
large RAID array, you might contact them not only for the system but also
support of this application. www.varesearch.com. They will probably try to
sell you Red Hat or S.u.S.E. but I think they can be persuaded to to
> 2) Work on the kernel-space NFS server seems to be moving along, albeit a
> bit slowly. Anyone have any comments about the performance of the
> user-space vs. the kernel-space NFS server programs?
Kernel NFS (what I call kNFS) has much better performance but there are
some issues. You must explicitly export each filesystem ... simply
exporting /usr does not mean you export any of the directories BELOW it.
If you want to also export /usr/share and others, you will need to
EXPLICITLY export them. User NFS (uNFS) exports subdirs with the parent.
Also, some filesystem types can NOT be exported.
> 3) I have been relatively happy with the performance of a Debian/i386
> system as a NFS server for a few home directories, shared filesystems,
> etc. How does the NFS subsystem scale at this time?
Talk to VA-Research they handle some rather large clients and would have a
better handle on this.
> 4) What are the biggest hardware constraints on system like this? I am
> investigating hardware RAID controllers. If the IO performance is
> maximized through hardware RAID, are these file serving processes
> (NFS/Samba/Netatalk) going to tend to be bound by CPU or memory
SAMBA is going to fork off a separate process for each user that is logged
in. You can discover a LOT of memory being taken up by SAMBA processes and
it can surprise you if you aren't aware of it up front. I would GUESS that
CPU and RAM will be your bottleneck ... the PCI bus will be too. Probably
RAM/PCI/CPU in that order depending on how much I/O you are doing.
> 5) What about coda? Is it stable enough to implement in a production
> system at the departmental level?
NO! You MIGHT get some performance boost out of something like reiserfs
but I would leave CODA alone for a while unless you are just using it with
a few in-house geeks so they can keep their laptops in sync. It is under
development and things are changing all the time. It is not very stable
yet ... still very much a moving target.