[Ltsp-discuss] NFS : can't get a request slot

[Martin H. <m.h...@gm...>]

Anselm Martin Hoffmeister schrieb:
> For excluding cabling & switch timing
> problems and the like, connecting the client to the server with a
> crossover cable seems a good idea. In the long run, switching those NICs
> for rtl8139 cheap 100MBit NICs (I currently buy them at less than 4 EUR)
> could be good idea as well.


Hi *

I have the same problem in some networks, installations in many other schools 
work fine. tcpdump or ethereal show a lot of UDP fragments when this happens.

Once I also had this problem with a rtl8139 card:
At the same switch, next to the server , the workstation woked fine, but
at the other side of the building, packages go (or don't go)
thru several switches & network-speeds : 100baseT--fiber-optic--100baseT
the same machine did not work anymore.

I did google and read about this issue today and what I'd like 2 try one day 
is to change the mount options on the client side. But to try this, I have to
build a new kernel+initrd image :-(

I'd like to try the following options:

rsize=1024, wsize=1024  ... to avoid fragmentation
tcp  		        ... to try NFS over TCP instead of UDP

another tunig option I found is to increase the Socket Input Queue:
(which might crash the machine after a while - you are warned)

echo 262144 > /proc/sys/net/core/rmem_default
echo 262144 > /proc/sys/net/core/rmem_max

(see: http://nfs.sourceforge.net/nfs-howto/performance.html )


Anselm: where do you buy rtl8139 for less than 4 EUR ?


Greetings, 
Martin Herweg


-- 
m.h...@gm...  GnuPG Key ID F058AF8E

Linux Hard- & Software für Schulen:
http://www.lug-marl.de/kmLinuxTSE.html