[OpenAFS] AFS performace question.

[Peter Popovics]

Hi there,

Running afs for the first time, and running out of the time assigned for me
to deploy afs in our department as generic  fileservice, I just realized the
very weak performance of AFS clients - at least in my installation. It seems
to be irrationaly slow, when reading from AFS filespace. I have done some
measures, by transfering some (~200Mbyte) general files of various size
(1K-6000K) to/from the afs filespace. It's not a "real measure", I didn't
have the patience to repeat them 3 times, but the numbers are reflecting
correctly my experiences...

1. using the AFS client on the same computer as the server: 
    (I686, Redhat Linux 7.0, 512M RAM)
      - Read: 1249 kbyte/sec 
      - Write:1274 kbyte/sec.
         ... This doesn't seems to be too bad...

2. using the AFS client on another Linux worksation
   (Redhat 6.2, 128 Mbyte of RAM)
    It was running the AFS server too, but the volume
    I was using was located on a different server.
      - Read: 20.15 kbyte/sec (!!!!!!!!)
      - Write: 167 kbyte/sec
       .... This is strange....

3. using the Windows AFS client on the 3rd workstation
    (I have done some tuning here: I have decreased the
    chunk size to 4k, statcache to 1024 entries, and the 
    cache size to 4096 kbyte... it's much more fast this way,
    even if it's the wrong way to solve this problem)
      - Read: 125 kbyte/sec (!!!!!!!)
      - Write: 733 kbyte/sec (this seems to be real)

4. some more general measures on the network:
    - scp between 2 linux boxes 824 kbyte/sec
    - nt copy between 2 NT boxes 820 kbyte/sec
    It's not too much on a 100M network, but still okay.

Measure 1-2 suggests that there is something wrong when the 
packets has to go out to the physical network. But measure 3,
and the positive effect of reducing the chunk size to a very low
value suggest, that there is something wrong with the servers
configuration. Unfortunatly I'm not too familiar with server debugging, and
the fileserver instance configuration - I assumed that the default 
setting should work fine, on a more ore less generic network
(measure 4). It's remarkable too, that the write supposed to be slower
than the read... but it isn't...

Using afsmonitor, I have noticed, that there are ~5% dropped packets,
and another value - rx_IgnoreAcked was growing considerable.

Some more information:
At the Linux boxes I have installed OpenAFS 1.0.2 from the RPMs.
Running the original build I did in December, I didn't experience these
problems - maybe I have never tried to use intensively.
On the NT (Win2k) box I'm running a "home-built" OpenAFS 1.0.2.

Probably it's some very basic problem... Please, let me know,
if you have any idea on this. Thanks in advance:

P.Peter