Bug 67625

From Bugzilla Helper:
User-Agent: Mozilla/5.0 Galeon/1.2.5 (X11; Linux i686; U;) Gecko/20020606

Description of problem:
I recently upgraded to the 2.4.18-5 kernel, and now I experience serious
performance problems with NFS v3 reads and/or writes, depending on the server.

With a Solaris 8 server machine (SunOS 5.8 Generic_108528-12 sun4u sparc
SUNW,Ultra-4) when I NFS mount with the default rsize (8192 now I believe) or
any rsize of 4096 or above, the read performance is very low (around
300Kbytes/sec). 
The 'nfsstat -c -r' command reports rapidly increasing RPC retransmissions: the
rate is 1.3 times the calls rate.
If I mount with rsize=2048 everything works out OK, and I have quite a good read
performance. The write performance is OK with various wsize options (I tested
1024, 2048, 4096 and 8192). This server has a 100Mbit connection. This server
worked flawlessly with rsize and wsize set to 8192 with kernel 2.4.18-4.

Now with an SGI server (IRIX64 6.5 10120105 IP25) the read performance is OK.
However the write performance is very very low (network traffic in the order of
32Kbytes per second) if I set wsize to 2048 or less. For 4096 or more the write
performance is OK. Looking at the network traffic with ethereal I remarked that
when wsize is 2048, each NFS write call is 512 bytes only (why not 2048?) and
that the reply always says "Commited: FILE_SYNC", no rpc retransmissions happen.
However with wsize of 4096 each NFS write is 4096 bytes and the answer always
says "Commited: UNSTABLE". I don't know if this scenario worked OK under
2.4.18-4. The server has a 10Mbit connection. The rsize setting has no notable
effect on the read performance, which is always good.

Version-Release number of selected component (if applicable): 2.4.18-5


How reproducible:
Always

Steps to Reproduce:
Scenario 1:

1. mount share from Solaris server with rsize=8192
2. copy a file from the server -> bad performance (~300 Kbytes/sec)
3. umount share
4. mount share from Solaris server with rsize=2048
5. copy a file from the server -> good performance (~677 Kbytes/sec)

Scenario 2:
1. mount share from SGI server with wsize=2048
2. copy file to server -> extermely bad performance (~30Kbytes/sec)
3. umount
4. mount share from SGI server with wsize=4096
3. copy file to server -> good performance (~466Kbytes/sec)

Actual Results:  Solaris server:
Bad read performance with lots of rpc retransmissions if rsize is more than 2048
(no badcalls seen at server). The amount of rpc retransmissions is about 1.3
times the amount of rpc calls.

IRIX server:
Extermely bad write performance if wsize is less than 4096. No rpc
retransmissions in this case (no badcalls reported on server either).

Expected Results:  Good read and write NFS performance.

Additional info:

My machine has a 100Mbit connection, but through a PCMCIA card (so max
performance is around the 12 Mbits per second I beleive).
My network card is a Xircom CreditCard Ethernet 10/100+ Modem 56 PCMCIA card
(CEM56-100) (driver xirc2ps_cs).