🚀 go-pugleaf

On 17 Nov 1996 22:47:07 GMT, James McOrmond <jam@server.jammys.net> wrote:
>David Jones (dej@igs.net) wrote:
>
>: In practice, if I set up NFS from my SparcStation to my Amiga 3000/16,
>: using 8K datagram sizes, I will get receiver overruns galore on the Amiga
>: whenever I read a large file.
>
>I found that anything larger than 1500bytes caused this, which just happens
>to be the max ethernet frame size.

  Yes, whenever you try to send packets larger than the MTU (maximum
transfer unit), they get fragmented. The fragments are then sent out,
usually in quick succession, and if a system can not accept back to
back packets, then a fragment will be lost, and the *WHOLE* packet
(all fragments) will get retransmitted.

  PC 8 bit ethernet cards were notorious for not accepting back to
back packets. I'm surprised that a Zorro II card would be so
underdesigned. Maybe it was the driver.
  TCP is aware of the MTU and can fragment things itself. It can also 
use MTU discovery to find the smallest MTU on the link (576 bytes is
the official minimum for IP. ATM has 48 byte frames. Can you see a
problem there?) 
 
  In the case of machines that seldom talk to each other on the same
logical "wire", such that they forget the ARP address of the other
machine, it gets even worse; the first fragment that goes out causes
an ARP request. The other fragments are dropped, since the arp cache
only keeps one outgoing packet in the queue. This occured regularly
some years ago at BNR. All NFS mounts went to 1024bytes. Performance
was not good :-(
  

-- 
   :!mcr!:            |  Network security consulting and 
   Michael Richardson |      contract programming
 WWW: <A HREF="http://www.sandelman.ottawa.on.ca/People/Michael_Richardson/Bio.html">mcr@sandelman.ottawa.on.ca</A>. PGP key available.