[rescue] AUI duplex [was Re: New SPARCStation 10 owner, needs help with some issues though.]

[Meelis Roos]

> > NFS-over-UDP can get so unreliable [...]
> 
> Oh, it can, it can indeed.  For much of my career I was co-admin of a
> network at a university where we used NFS-over-UDP heavily.  I've seen
> plenty of rather interesting failure modes.  (I think the most
> interesting was one where NFS reads of files would hang when the file
> size fell into one of certain small ranges.  Turned out to be a minimum
> packet size issue: one host was generating packets small enough that
> another host would drop them as runts.  UDP overhead meant this
> happened only with last frags, hence the odd file size dependency.
> Remounting with rsize=1024 made the symptom go away, as it meant
> IP-layer fragmentation never happened.  I never found a reference
> authoritative enough to tell me which host was de-jure at fault.)

Reminds me a case of NFS trouble where rsize=1024 solved the problem. 
That was with serving NFS from fast server (486) to slow PC-s with ISA 
NE2000 compatible NICs. With default rsize of 8192 or 4096, the packet 
was fragmented and fragments were sent as quickly as the network 
allowed. However there was no buffer space availble in the NICs to store 
all the fragments at once and it depended on the host to read them out 
quikly. Unfortunatelty the slow PC could not receive the packets fast 
enough and the last fragments were always dropped. Of course the same 
happened to the retransmits...

-- 
Meelis Roos (mroos at linux.ee)