[clue-tech] MAC NFS?

Nate Duehr nate at natetech.com
Fri Sep 14 00:13:08 MDT 2007


On Sep 13, 2007, at 11:26 AM, David L. Willson wrote:

> Anyone aware of a way to skip the TCP/IP layer when NFS mounting?   
> Both
> boxes are on the same segment, and I want maximum throughput, lowest
> latency, lowest CPU demand, for free, with a shared network filesystem
> (hence NFS).  This storage will be used to host vmware files for 2-4
> Linux VMware Servers.

NFS by default is UDP, not TCP.  Perhaps you have it configured wrong  
if you're seeing TCP packets, and are using NFS over TCP, a  
"relatively new" feature.

Additionally, NFS uses DNS to find where it's going unless you're  
using IP's in the NFS mounts.  Static entries in /etc/hosts and the  
appropriate /etc/nsswitch.conf configuration (most distros have the  
correct configuration here already, using the hosts file first),  
could speed things up a little.

Mostly it sounds like you need a text on NFS tuning and/or drop NFS  
altogether.   The Linux NFS-HOWTO has a lot of information.

http://nfs.sourceforge.net/nfs-howto/

However...

You say you're hosting VMWare files.  Are you talking about the  
VMWare files holding the entire virtual machine?  How do they boot?   
How much RAM do the servers have, can they cache the entire file as  
well as run from it at the same time?  (NFS does file caching in RAM,  
of course.)  Your bottleneck may not be NFS at all.

I think VMWare themselves make solutions for this, in their "server"  
products... they're more expensive than the freebie versions, but I  
don't think you'd need to mess with NFS, etc.  You just load up  
VMWare and go... using their "management" tools.  Their apps manage  
pushing the images around to the hardware.

Free is great, but if you're setting up something that must be  
virtualized for a production environment, having at least a small  
budget seems reasonable.  I understand if it's not, though.

But look at it carefully to see if NFS is REALLY the bottleneck.  You  
mention that the two boxes are on the same segment -- is the segment  
busy with other traffic?  Could a dedicated GigE port be put in each  
machine as a "backend" poor man's Storage Area Network?  Are the  
disks in the NFS server slow?  Are they at least RAID 0/1 for read  
speed enhancement?  Etc etc etc...

How have you quantified that NFS is the "slowness" culprit?  In other  
words, as an engineer friend of mine likes to say, "Are you a  
thinkin' man, or are you a knowin' man?"  :-)

Well, anyway... the NFS-HOWTO has some good tips on tuning.  Really  
good tips.  When you see things like, "This test will take at least  
three hours..." someone did their homework and wrote a nice  
document.  :-)  Check it out.

--
Nate Duehr
nate at natetech.com






More information about the clue-tech mailing list