Re: AmigaNet warning messages

[Stefan Hensen <hensen%wpos4.physik.uni-wuppertal.de@localhost>]

On Fri, 27 Aug 1999, Ignatios Souvatzis wrote:

> [About the ed0: warning - receiver ring buffer overrun messages]

> NFS, with standard parameters, likes to send 8 kBytes of data + rpc headers
> at once. Such a packet gets split up into 6 fragments on Ethernet. When the
> receiver chip can't buffer them all, and the receiving cpu isn't fast enough
> to read them off the chip fast enough, some get lost, or overwrite earlier
> packets, depending on the Ethernet chipset used. The latter is what happens
> to the Amiga ed users.
>
> Set the "nfs read size" and "nfs write size" (nfs_mount: -r / -w options) to
> 1 kByte on the client machine.

Setting the "nfs read size" to 1 kByte indeed solves the problem. Even
setting it to 4 kByte works fine, making file transfers about 33% faster.
As already mentioned in my previous mail, writing to NFS volumes causes no
problems, so the "nfs write size" can be left at the default value.

> FTP, rcp etc.  don't stuffer from this problem because they operate over
> TCP, which knows about the packet size on (e.g.) Ethernet and only sends 
> (on average) 2 packets before waiting for an ACK packet from the peer.

I already thought that the use of UDP instead of TCP could be part of the
problem, but did not understand why this only affects the client operations
and not the server operations. However, with limited buffer size / too slow
CPU being the real reason, this seems to become clear, simply meaning that
the ethernet chip is probably always fast enough to handle all data it gets
from the CPU, but not vice versa.

Thanks for the information.

Regards,

Stefan Hensen

----------------------------------------------------------
Stefan Hensen
e-mail: hensen%wpos4.physik.uni-wuppertal.de@localhost