James Carlson wrote, on 11-01-10 15:38:
> Is there any chance you could make the raw tcpdump files (rather than
> the decoded text output) available?  It's much harder to debug things
> when information has been subtracted.
> 
> Based on what little is available here, I'm seeing what seems to be an
> RTT that jumps around a lot, from a few milliseconds on up to 150
> milliseconds, and causing TCP to misbehave.  It's likely that this is a
> kernel issue that needs to be investigated.

I created an new tcpdump of the issue and attached it in the mail.

http://filebin.ca/zmmnws/tcpdump01.tar.gz

I changed the sammy servers pppoe connection to an normal eth5
interface to an ciso router that is providing a 50/50mbit fiber
interface. So the upload speed of the sammy server should be very
high. But is only 30KB when uploading to my mobile umts systems. When
uploading form an other server to the umts systems the upload speed is
the max of that server (100kB) so the umts systems should not be the
problem here.

We opened the tcmpdumps with wireshark and did an analyse > export
info and it showed a lot of "TCP Fast retransmissions" compared to the
the other tcmpdump of the good download from the ashley server.

I also tried the below kernels, but it did not change anything.

linux-image-686 | 2.6.30
linux-image-686 | 2.6.31
linux-image-686 | 2.6.32
linux-image-amd64 | 2.6.30
linux-image-amd64 | 2.6.32

If anybody can have a look at the tcp dumps or maybe have other
suggestion to get more debug info?

Best regards,

Jelle