NFSD - How to determine how many nfsd daemons to run.

[LV2DIVE]

Environment:
Server:
HP c8000 w/ Compellent drive as SAN/DATA server
HP-UX 11.11

Total number of clients are 50 machines made up of these types of machines:

Qty - Type of machine
35 - HP c8000 HP-UX 11.11
5 - HP c3750 HP-UX 11.11
9 - IBM p275 AIX 5.1
1 - SGI Octane2 IRIX 6.5

What is a good starting point with six (6) mount points on every client to the same server?

[Steven E. Protter]

Shalom,

Almost all of the time, I found that the default configuraion of 16 nfsd processes works well.

You should periodically run nfsstat and look at the data. If you find you are often at more than 16 connections then you may improve performance slightly by increasing the number of startup nfsd processes. NFS will however open new processes as needed with little performance penalty.

SEP

[Dave Olker]

Hi,

The most important question in sizing the number of nfsds is what transport protocol are the clients using to mount the filesystems. Are these clients using UDP or TCP to access the NFS filesystems? If you're not sure, you can try issuing the following command on the server and post the output here:

# nfsstat -s

Regards,

Dave

[LV2DIVE]

# nfsstat -s

Server rpc:
Connection oriented:
calls badcalls nullrecv
5863089 0 0
badlen xdrcall dupchecks
0 0 85292
dupreqs
0
Connectionless oriented:
calls badcalls nullrecv
32226 0 0
badlen xdrcall dupchecks
0 0 7475
dupreqs
0

Server nfs:
calls badcalls
5898043 0
Version 2: (32577 calls)
null getattr setattr
10 0% 4596 14% 0 0%
root lookup readlink
0 0% 2861 8% 5871 18%
read wrcache write
19177 58% 0 0% 0 0%
create remove rename
0 0% 0 0% 0 0%
link symlink mkdir
0 0% 0 0% 0 0%
rmdir readdir statfs
0 0% 18 0% 44 0%
Version 3: (5865417 calls)
null getattr setattr
106 0% 335953 5% 2224 0%
lookup access readlink
380447 6% 88943 1% 1863 0%
read write create
4905976 83% 87699 1% 1010 0%
mkdir symlink mknod
15 0% 0 0% 0 0%
remove rmdir rename
1129 0% 0 0% 337 0%
link readdir readdir+
353 0% 320 0% 44911 0%
fsstat fsinfo pathconf
396 0% 106 0% 1 0%
commit
13628 0%

[Dave Olker · Accepted solution]

> Server rpc:
> Connection oriented:
> calls
> 5863089
>
> Connectionless oriented:
> calls
> 32226

This server is doing predominantly TCP, so the number of running nfsds on the system is meaningless. The NUM_NFSD parameter only sizes the UDP daemons.

The way TCP works is as follows:

Each of your 50 clients will open a separate TCP connection to the server. The server will create a pool of up to 10 threads (by default) dedicated to each TCP connection. This means you could potentially have up to 500 nfsktctpd threads on your server at any given time.

However, since the server creates and destroys threads on an as-needed basis, chances are there are probably between 50 and 100 threads on the system, based on the amount of traffic coming from the clients. You can use Glance to verify the number of running nfsktcpd threads if you're curious.

Is there a reason you're wanting to change the server configuration or were you just curious how it all works?

Regards,

Dave

[Rasheed Tamton]

If you are really worried about performance of the NFS server it might be a good option to look for autofs (previously automounter). It might be availabe for the other OSs (AIX/IRIX) also.

Otherwise, you can even look for Samba. A well tuned Samba server can be an alternate option for NFS and it comes bundled with the latest hp-ux versions.

NFS V3 uses TCP.

Rgds,
Rasheed Tamton.

[LV2DIVE]

I was only trying to get the most out of the server. I did not realize how ever that nfsd are only for udp and not tcp. Just looking for the optimal settings to have happier users.

Thank you all both for you input.

[LV2DIVE]

Thanks to everyone for your input.