Very Computer

Configuration:
--------------

Machine  OS     Version    NFS    
-------  ---    -------    ------
A1       AIX    3.2.51     Server    
A2       AIX    3.2.51     Client
L1       Linux  1.3.64     Client

Mount output for nfs partition for each machine
------------------------------------------------
A1: /dev/fslv01 /home      jfs    Feb 06 01:28 rw,log=/dev/loglv02
A2: A1 /home    /home      nfs    Mar 19 13:50 rw,hard,bg,intr
L1: A1:/home on /home type nfs    (rw,hard,intr,addr=XX.XX.XX.XX)
    where xx.xx.xx.xx is the correct local address of the server.

        It is not a network bandwidth problem, as ftp and other
networking apps on the linux box are at full speed - only nfs is
bottle-necked.  And A2 and L1 machines are both on the same ethernet
subnet as A1, so the network cabling/routing to the server is not a
problem.

Performance:
------------
The following gives the performance of creating a 1MB
file (using dd) on the /home file partition from the A2 and L1 nfs
clients (using a much larger file as the input file to dd).  As you
can see, it does much worse for the Linux client L1!  I would be happy
if the linux nfs client could do as well as the AIX nfs client A2...

Command: time dd if=/home/LargeFile of=/home/tstMM count=2k
         where MM is the machine name
Results:
A1:   1.33 seconds   (local file system, so no nfs overhead)
A2:  11.98 seconds   (nfs client of A1)
L1: 157.94 seconds   (nfs client of A1)

Questions:
---------

1. Any idea of nfs client parameters we can tweak on the linux box L1
to get acceptable performance?

2. When the L1 is nfs writing to A1:/home, you can actually hear the
hard disk sort of thrashing on A1 - maybe L1 is writing one byte at a
time?  How can I check?

3. If it is the AIX version of nfs that is screwy (optimized to work well only for other nfs machines), can we tweak the linux nfs version to act like an aix nfs client? (the linux box is only nfs mounting the aix box).

4. Maybe this can be fixed by manipulating the datagram wsize and
rsize parameters of mount (pg. 172, NAG) - if so, does anyone know
what size they should be for working with AIX NFS servers?

thank you,
Brad L. Miller

Hi Brad,

   >I am having severe performance problems using linux as a nfs
   >client with an AIX nfs server.  The configuration is briefly described,
   >then the performance stats, and then a couple of questions.

I had the same trouble with an IRIX NFS server and Linux client. For
me, the key variable was rsize,wsize. I've got rsize=8192,wsize=8192
in my Linux box's /etc/fstab now.  Linux defaults to a 1024 byte
read/write size. This change boosted my NFS performance in simple cp's
and dd's by 8-fold.  It got me to within 1.5-2x the performance of my
commercial (IRIX and Solaris) UNIX NFS clients, which I can live with.

I still take a fairly severe (4-fold) hit in jobs that do a lot of
small read/writes (such as an 'ld' of a lot of object code files,
during a 'make').  I'm running 1.2.13. I was told that 1.3.x kernels
did/would have additional NFS client improvements but haven't the time
to play with the experimental kernels.

Yours,
Sean Eddy
--
- Sean Eddy
- Dept. of Genetics, Washington University School of Medicine