NFS Limitation?

[Apollolin]

Dear all,

Thanks in advance for your reply and sharing.

Suppose there are two envrionment: Server_A with Client_A1~Client_A50; Server_B with Client_B1~Client_B50. Each server shares some directories, say /Share_A/ and /Share_B/, respectively, to its clients through NFS (assume mount point is also /Share_A/ and /Share_B/ on the clients).

For some reason, Client_As have to access /Share_B/For_Client_A/ and Client_Bs have to access /Share_A/For_Client_B. Since each server already has its own 50 clients, to avoid heavy loading, I use NFS to mount Server_B:/Share_B/For_Client_A on /Share_A/Mount_from_B, and Server_A:/Share_A/For_Client_B on /Share_B/Mount_from_A. By doing this, I hope I can access some data on Server_B on Client_A while Client_A does not directly communicate with Server_B, and which may avoid the possible heavier loading on both servers.

After the experiment, I can access the required content of the other server in the current server. However, it does not apply to the clients. For example, on Client_A1, I can see /Share_A/Mount_from_B directory, but there's nothing in the directory (ls -l => total 0.) while the content can be seen on Server_A console.

I try looking up related information on the Internet, but there's no exact anwser to my question; I believe I did not choose precise keyword while searching, but I don't really know the terminology for my case (Cross NFS mount? Overlay?). I used to hear a restriction posed on NFS that we can not export a directory which has already been mounted using NFS. Is it true? Or is there different mounting scheme to deal with such condition?

I'd really like to know if there's a better way to manage this problem. Adding hardware is the last resort, oops. If such kind of NFS mount through two machines is not possible, maybe I'll try to dynamically mount other server whenever access is required; but I am not sure if this method is practical.

Millions thanks for your suggestion and sharing!

P.S. All the workstation operates in HP-UX B.11.11 U 9000/785 environemnt.

Apollo

[Steven E. Protter]

Shalom,

What you have attempted is apparently a NFS load balancing scheme.

Possible alternatives:
1) Locate the data on an EMC like device that handles load differently than a Unix Server.
2) Set up an rsync of changed files between two servers and use DNS round robin to assign connect names at random.

In these or other schemes you must make sure user and permissions are the same on both systems or you will run into access problems.

SEP

[OldSchool]

so........

serverA:/clientA --mounts---> serverB:/shareA

serverA:/clientA is exported.

and clientA ---mounts---> serverA:/clientA

If I recall correctly, you will never see the contents of serverB:/shareA on clientA, only the mount point directory. As you noted, you can't export the NFS-mounted filesystem. Doing so exports the *mount point* which is the empty directory where the filesystem is mounted.

This is noted in the man pages:
"A server can also be a client with respect to file systems it has mounted over the network; however, its clients cannot directly access those file systems. If a client attempts to mount a file system for which the server is an NFS client, the server returns with errno set to [EREMOTE]. The client must mount the file system directly from the server on which the file system resides."

As to overhead, if it worked, it would create more than you're attempting to save. Consider this: A write to share mounted on the client has to got thru NFS to serverA. NFS on serverA would have to figure out it doesn't have the real location and pass the write to serverB. ServerB would have to do the write, pass the "i've written it" back to servera, which it turn has to pass the "iv'e written it" back to the client.

As far as I know, this is a self-defense mechanism to prevent accidental creation of circular references. Imagine what would happen in the above if you were to attempt to add:

serverB:/shareA ---mounts---> serverA:/clientA

By the way, 50 clients is a fairly "light" load for NFS. I believe I'd skip automounter and just mount what you need where you need it.

If you've issues with performance, consider starting more biod's

[Dave Olker]

Hi Apollo,

As OldSchool mentioned, this is not allowed by the NFS protocol as it violates one of the primary security tenets, which is that the NFS server decides which clients are allowed access to the files it controls.

Is there any reason why the 100 clients can't just mount filesystems from both servers?

Regards,

Dave

[Apollolin]

Dears,

Thanks so much for your reply and advice.

Actually, the client is measurement equipment which requires some specifications and programs on the server and saves the measurement data back to server for further analysis. Since the data and specifications and programs are critical, we build a twin-server, say Server_A2, as Server_A1's image. Therefore, 50 clients and Server_A2 and other systems (such as EDA tools, failure anaylisys, etc) are continually accessing Server_A1, resulting in loading issue. Server_B1, Server_B2, and Client_Bs are in another site. For some reason, Client_Bs/Clinet_As have to use the specifications and programs on Server_A/Server_B and then store the data back. Since current loading of Server_A and Server_B is not low, we are not sure if adding another 50 clients to both servers will impact the performance or not.

Unfortunately, since NFS has such limitation and adding hardware is not under our control, it seems we have to increase more clients on both servers to meet the requirement. Maybe I'll try dynamic NFS mount/unmount whenever access is demanded or restrict the required access to server in other site only some specific local clients to avoid the possible loading rage.

What do you think?

Thanks again for your comment and sharing.

Apollo

[OldSchool]

"Since current loading of Server_A and Server_B is not low, we are not sure if adding another 50 clients to both servers will impact the performance or not."

As I noted previously, 50 clients is a light load as far as NFS, so what make you think think that NFS is the issue?