Bug#364607: reproducible in etch?

[Arthur de Jong <arthur@west.nl>]

On Thu, 2008-12-04 at 23:49 +0100, Moritz Muehlenhoff wrote:
> Does this error still occur with more recent kernel versions?

I've just reran the tests with the same hardware. However the client
machines are tracking testing. The server is running Debian/etch with
2.6.18-6-686 and nfs-common 1:1.0.10-6+etch.1, the clients are running:

 host   OS              kernel        nfs-common       HW
 ------ --------------- ------------- ---------------- -------------------
 host1  Debian/testing  2.6.26-1-686  1:1.1.2-6lenny1  Dell Optiplex SX280
 host2  Debian/testing  2.6.26-1-686  1:1.1.2-6lenny1  Dell Optiplex SX280
 host3  Debian/testing  2.6.26-1-686  1:1.1.2-6lenny1  Dell Optiplex SX280
 host4  Debian/testing  2.6.26-1-686  1:1.1.2-6lenny1  Dell Optiplex SX270
 host5  Debian/testing  2.6.26-1-686  1:1.1.2-6lenny1  Dell Optiplex 745
 ------ --------------- ------------- ---------------- -------------------

 host  speed discard   nfs  nfs write  nfs read
      Mbit/s    MB/s  prot       MB/s      MB/s
 ----- ----- ------- ----- ---------- ---------
 host1  1000     104   udp         25         7
                       tcp         27        11
 host2   100      12   udp         10         1
                       tcp         10        10
 host3   100      11   udp         10         3
                       tcp          7         5
 host4  1000     112   udp         21         8
                       tcp         22         7
 host5  1000     118   udp         35        10
                       tcp         35         9
 ----- ----- ------- ----- ---------- ---------

The tests:
 - discard:
   # dd if=/dev/zero bs=1024k count=1024 | nc -q 0 server discard
 - create /tmp/tmpfile:
   # dd if=/dev/urandom of=/tmp/tmpfile bs=1M count=1024
 - mount /mnt with either of:
   # mount server:/share /mnt -t nfs -o rw,udp
   # mount server:/share /mnt -t nfs -o rw,tcp
 - write and read tests:
   # dd if=/tmp/tmpfile of=/mnt/tmpfile bs=8k
   # dd if=/mnt/tmpfile of=/dev/null bs=8k

The network was otherwise in normal (not too heavy) use so the results
may be a little off (the NFS performance over TCP on host3 surprised me
a little but I couldn't really get any better numbers than this).

I would say the problem is still there. The machines that run on 100Mbit
have slower read performance when on UDP than on TCP. If you want I can
also test different client/server combinations but that will be in the
new year and these test cost quite some time to execute.

-- 
-- arthur de jong - arthur@west.nl - west consulting b.v. --