I tried to copy 15GByte over NFS: source single Pentium 400 , plain Suse 7.0 this networkcard on both sides (100 MBit, duplex, nway switch )
This sounds well-known to me. We are also using a couple of rtl8139-cards, and copying larger files over NFS leads very fast to a stopped network. The only workaround I found is to start ntop on the NFS server for about a second. But it's annoying and not secure to start ntop every time the network hangs this way.
Btw, I also found, that the userspace NFS daemon, shipped with SuSE [6.4] hangs less than the knfsd (SuSE-linux-2.2.16). Another solution we tried was to reduce the network speed back to 10MBit, and everything worked fine (but this option is out of question).
If there are known solutions and patches, please PM to me.
Yes, there are: build a new kernel and set 'CPU is to slow to handle full network bandwith' to yes this slows down your client, but makes this error disapear. The error is a known isue with fast nfs servers: the server pushes out nfs packets faster than the client is able processing them. After some time there maybe a buffer overrun at the clients side. Some packets get lost. The client requests resending them. Again they arrive to fast to be processed... Since the knfsd *is* faster than the user space nfsd it's more likely appearing with knfsd. There are two solutions: 1st chance is to slow down your server. Starting ntop does it temporarily or 2nd make shure no more packets arrive as your client is able to process. You do it changing the mentioned kernel flag. Than building a new kernel for your client. Thomas