Hello!
> > I have tested this with Linux 2.2.12, 2.2.14, 2.2.16, 2.2.17 and
> > 2.4.0-test6 kernels on the sender, and Solaris 2.6 and 2.7 for the
> > receiver, and the results are the same.
....
> > From: Vern Paxson <vern@xxxxxxxxxx>
...
> > Well, the problem is that the sender is miscomputing its RTO, setting it to
> > a value that's much too low (according to the standard).
I hope Paxson interpreted the problem incorrectly. 8)
(following his own researchs, which are not quite right and do not
apply to linux at all)
I think real reason is that Linux had bug estimating rtt using as sample
the last acked segment, so that delack bias was not taken into account
at all.
Could you check this with 2.2?
In fact the fix is very simple: in tcp_clean_rtx_queue() use __s32
for seq_rtt, intialize it to something negative and update only
when it is negative, so that only the first segemnt is used for sampling.
Can you make it yourself? (I do not have any 2.2 tree right now.)
Alexey
|