Hello !
The solution is much simpler than I expected. Paul McDowell pointed
me at the possibility to make /home itself a CDSL and mount the
user directories on a per member basis (on nominal /home/user dirs).
So the mount service is not only highly available (without actually
using CAA), but also has a very high bandwidth since nothing has to
be pumped through Memory Channel.
But two minor disadvantages of this solution should be mentioned here:
1. We see ALL member mounts when typing 'mount', since the cluster sees
also CDSLs through the CFS.
2. The 'pwd' command gives back the actual mount point, not /home/user.
But 'echo $PWD' or 'echo $cwd' ((t)csh) do show the nominal mount
point. This problem may also affects Fortran inquire() statements or
C getcwd() functions.
Thanks to Paul for his quick reply !
----------------------
Paul McDowells answer:
We had a similar problem to this when first testing V5 with NFS mounts. We
found that if a cluster member crashes that was serving the NFS mount to the
other members then the mount becomes unavailable as it is NOT relocated by
the CFSMGR!
To get round this we used the context-dependent symbolic link (CDSL) utility
which allows NFS clients to be mounted on a member specific mount point.
For example instead of just mounting, say, a home file system on one node of
the cluster on /home and then having it served to the other members via
memory channel, creating a CDSL for /home means that EACH node can
independently mount the client. So if its a two node cluster the NFS file
system is actually mounted on /cluster/members/member1/home and
/cluster/members/member2/home improving both performance and availability
but the user still access the filesystem via /home so its bascially
transparent.
Check the man page on mkcdsl for a better explanation.
=========================================================================
Original message:
>>
>> TruCluster 5.1 has the restriction that NFS mounts (as client) are
>> done on a per member basis, and this single member then exports the
>> mounted filesystems clusterwide. As we have several users with high
>> bandwidth traffic (SUN NFS server), I decided to distribute user
>> directory mounts among the members of the cluster to balance network
>> loads.
>> But this means that some filesystems will not be available if a
>> member is shut down. Autofs provides a way of failover as a CAA
>> service, but it also mounts a complete map on only one member.
>> My idea is to run several autofs daemons, each with it's (disjoint
>> from others) auto_map files, which serves a specific group of
>> users, and has a set of preferred servers for failover.
>> But I don't know how to register several autofs daemons in CAA with
>> their respective maps. Must there be alias names for each autofs
>> (symbolic links, as autofs_1,...) to provide different names
>> for the CAA service ? Is it even possible to rename the autofs daemon ?
>> Is the whole idea feasible ?
--
Dr. Udo Grabowski email: udo.grabowski_at_imk.fzk.de
Institut f. Meteorologie und Klimaforschung II, Forschungszentrum Karslruhe
Postfach 3640, D-76021 Karlsruhe, Germany Tel: (+49) 7247 82-6026
http://www.fzk.de/imk/imk2/ame/grabowski/ Fax: " -6141
Received on Fri Jul 27 2001 - 13:23:15 NZST