RE: [lwip-devel] TCP ack timing

[K.J. Mansley]

On Fri, 2004-04-30 at 16:17, Mountifield, Tony wrote:
> > Anyway, I think that even with that change the original 
> > poster would see the same problem, as the max segment size
> > must be less than half the receive window for the algorithm
> > to work.
> 
> The strange thing I'm finding is that by taking the MSS back down to
> smaller values, my lwip system crashes, so I'm wondering whether there
> is some problem with that algorithm on my system.
> 
> The window size is still currently 2048.
> 
> If I make the MSS 128, 256 or 512, and then send the data stream to
> it, it crashes very quickly. If I make the MSS 1000 it crashes after a
> while. If I make the MSS 1024 it appears to be ok.

Sounds like a bug somewhere else - possibly the code that sends the
delayed acknowledgement!  The fact that it only happens when your MSS is
< half the receive window suggests these two factors are related, but it
looks like the large MSS (in relation to the RCV_WND) is the reason for
the bursty behaviour that you're seeing.

I checked the FreeBSD code, and they have a similar approach (send when
segments >= 2, rather than checking the amount of outstanding
unacknowledged data) to the current lwIP, so I think that aspect is
fine.  We should probably add a comment to lwipopts to point out the
effects of setting TCP_MSS to be large though.

Kieran