I have a NFS share which is used to store email accounts. So, you can imagine the number of read & write operation over NFS. 80% write operations!

The NFS server runs the following hardware & software:
Supermicro 2u, RAID-10 (Sata RAID, 3ware controller), 1.5TB NFS-Share, 8GB RAM DDR2, Pentium D 2.6GHz processor (looks like a desktop processor to me on server board, perhaps!), baseboard - Supermicro PDSML(not details available)

OS - CentOS 5.2 64bit, NFSv3 operation (installed from nfs-utils-1.0.9-33), current kernel 2.6.18-194.3.1.el5

I am facing a big performance issue with NFS.. the load avg on the nfs servers are always between 10 and 30. I had to increase the number of nfs daemons to 12 since its shared across 7 clients and most nfsd go in ‘D’ state. If I keep default, then mail server ends up with high mail queue and lots of lock failures. We also need to maintain nfslock since file locking is important to us.

client side mount options used are:
hard,intr,lock,auto,sync,tcp,rsize=32768,wsize=32768

server side options are:
$share 172.16.1.0/24(rw,sync,root_squash,no_wdelay)

Output of nfstat
Server nfs v3:
null getattr setattr lookup access readlink
140 0% 494082 1% 66987 0% 358388 1% 420319 1% 0 0%
read write create mkdir symlink mknod
1805704 6% 24135656 88% 23579 0% 1 0% 0 0% 0 0%
remove rmdir rename link readdir readdirplus
21374 0% 0 0% 24297 0% 21362 0% 0 0% 11 0%
fsstat fsinfo pathconf commit
221 0% 8 0% 0 0% 0 0%

Output of mpstat:
08:15:35 AM CPU %user %nice %sys %iowait %irq %soft %steal %idle intr/s
08:15:40 AM all 15.02 0.00 9.01 24.42 0.20 0.50 0.00 50.85 1836.20
08:15:45 AM all 13.89 0.00 7.09 26.47 0.00 0.20 0.00 52.35 1313.20
Average: all 14.45 0.00 8.05 25.45 0.10 0.35 0.00 51.60 1574.70

However, I have an NFS server from the same pool that functions remarkably. Its the same config as the rest. Its nfs write operations are 84% and iowait is 0.05%. The only thing I can see about this server is that the OS was able to detect the family of the processor exactly (i.e. All processor Flags | use dmidecode -t processor), while for the rest it couldnt. Let me tell you that processor is exactly the same and so is the OS.

Is there anything I can do to improve NFS performance?

Thanks!

Thank you for the article, which I found helpful.

In the section- NFS Client Configuration

NFS volumes can be mounted by root directly from the command line. For example
..
..
..
..

For each options menctioned in /etc/fstab file check the man pages
of fstab.”Click here for manpage.” - There is no link provided here.

Regards,
Veena

apt-get is a thin, low-level command-line interface to libapt. aptitude is a user-friendly front-end. You should not normally use apt-get unless you really know what you’re doing. In particular, aptitude has better conflict resolution, knows how to pull in Recommends, and can automatically remove packages that are no longer needed (and were only installed as dependencies of other packages).

I know the word “apt-get” is deeply ingrained in common knowledge, but really, aptitude is what most people should be using these days.

Thanks for your valuable information and may be i am still under learning stage please check my comments as follows

* The kernel-mode server is _more_ featureful than the userspace server (and 2.2.13 isn’t really a “recent” kernel by any standard, the 2.2 series was abandoned like five years ago). For one, it supports NFSv4, which the article for some reason doesn’t cover at all.

I am using debian stable version so i am using nfs package coming with this.
* The default rsize and wsize have not been 1024 in ages; I believe they’re 32768 now for NFSv3, but I haven’t checked it. Better to leave the field alone.

I have menctioned about the NFSv3 they have to use 32768 under performance.I am not sure every one is using NFSv3 so this is general writeup

This is off topic, however I was wondering why you recommend using aptitude over apt-get?

I’d always thought that aptitude was just a frontend to apt-get and have never used it (when I came to debian years ago from RedHat/Slackware apt-get seemed like the best thing in the world) so I was wondering if I’m missing something important?

Cheers, GSandie

As the Debian NFS maintainer, I have to say I’m a bit disappointed at the quality of this article; there are so many things that are just plain wrong. I can’t address them all in too great detail, but a few points are (and I hope the website doesn’t mess up my list ):

* You do not install things with apt-get unless you really know what you’re doing; use aptitude.
* The kernel-mode server is _more_ featureful than the userspace server (and 2.2.13 isn’t really a “recent” kernel by any standard, the 2.2 series was abandoned like five years ago). For one, it supports NFSv4, which the article for some reason doesn’t cover at all.
* The default rsize and wsize have not been 1024 in ages; I believe they’re 32768 now for NFSv3, but I haven’t checked it. Better to leave the field alone.
* -o soft or -o hard has absolutely nothing to do with symlinks vs. hardlinks. Soft mounts do not cause data corruption just by themselves; it has to be combined with some kind of outage (like a network outage).
* If you have any kind of network where you could expect any sort of packet loss, you most likely want TCP. In fact, NFSv4 only lets you use TCP, which is a good thing.
* NFS supports Kerberos authentication in addition to trusting IPs; it’s not so horribly insecure as you claim.
* Lots of memory _is_ helpful on both clients and servers. Both can cache, which helps hit the disks less.

I appreciate people writing articles on difficult topics, but please, make a better job of checking your facts first instead of copying them blindly from man pages or old versions of package descriptions.