|
Most excellent. We will proceed and let this list know the results.
Thanks,
Joe
Joseph Mack NA3T wrote:
On Tue, 7 Mar 2006, Joseph L. Kaiser wrote:
Hi,
I have been tasked to mount a read-only NFS mounted software area to
500+ nodes. I wanted to do this with NFS and LVS, but after reading
all the howtos with regard to NFS and LVS and reading all the email
with regard to this in the archives (twice), it seems clear to me
that this doesn't work.
(I haven't done any of this, I've just talked to people - so my
knowledge is only theoretical).
(ro) is a different kettle of fish. If you use identical disks (or
identical geometry) and make the disks bitwise identical (eg made with
dd), then clients will always get the same filehandle for the same
file, no matter which realserver they connect to. I assume you won't
be able to use RAID etc, just single disks.
For disk failure, make sure you have a few dd copies spare. If you
have to update files, then you'll have to do it bitwise. I don't have
any ideas on how to do this off the top of my head. Even though you
don't have the failover problem anymore, you still can't do scheduled
maintenance, which is much the same thing.
However, I have a boss, and he wanted me to ask if turning off
no-attribute caching (noac) would help in the reliability of this
service.
If the disks are (ro) then the attributes on the server will never
change, in which case you want the client to cache the attributes
forever.
He has seen with another NFS mounted filesystem that using the "noac"
turns off caching and clients that sometimes get "Stale NFS file
handle" will reread the file and succeed. So my question is:
Is "sometimes" an acceptable spec? Will it now pass the tests in Sect
12.32.9. "stale file handles"
1. Has anyone seen this behavior with "noac"?
2. Does it make sense to try turning off "noac" and testing? Will
the filehandle problem still persist and make this just a dumb thing
to do?
It will take you 10mins to do the tests in sect 12.32.9
(be sure to let us know the results)
Joe
|
