> On 1 Apr 1996, Christoph Lameter wrote:
> >
> > I think I isolated the problem. Alan changed the ACK handling in 1.3.73. Ever since then
> > we have the performance problems because Packets are resent and resent.
> >
> > Look at the 1.3.73 patch at the changes in linux/net/ipv4. Sadly I dont have the
> > knowlege to fix it.
>
> Don't blame Alan, it was probably me. I've been working on TCP issues,
> but I have only fast interfaces to play with (loopback and normal
> ethenet, soon fast ethernet). Those two are faster these days, but it
> could have impacted slow interfaces negatively (I don't see how, right
> now, but I might be missing something).
>
> How isolated have you gotten it? Can you check if it's the tcp.c or
> tcp_input.c changes? (if it's the retransmit changes, _then_ you can
> blame Alan ;-)
For what it might be worth, runnung 1.3.82 and watching tcpdump at
each end we see:
Using ftp over ppp through a server to a Sun this linux box seems to fail
to ack, there's a multi-second pause and the Sun retransmits, we ack and
proceed a little further until linux fails to ack and the process repeats.
The ftp session tends to degenerate further as time passes in that the Sun
(4.1.3) seems to not retransmit as often.
Using ftp over slip there are pauses (shorter) where the linux tcpdump
reports truncated-ip xxx bytes missing.
I have no idea as to which xxx.c might be the problem. One might suspect
several problems since ppp and slip behave differently.
Might be good if Chuck Hedrick would look at this. He's fixed a lot of
ip stacks over the years...
--joe