Extremely slow web page loading to only 1 site ?
Linuxguy123
linuxguy123 at gmail.com
Wed Feb 9 21:53:31 UTC 2011
On Tue, 2010-02-09 at 11:25 +0800, Ed Greshko wrote:
> Bruno Wolff III wrote:
> > On Tue, Feb 09, 2010 at 08:21:33 +0800,
> > Ed Greshko <Ed.Greshko at greshko.com> wrote:
> >
> >> In the back of my mind, from a few years ago, I remember there an issue
> >> with the linux TCP/IP stack and some "mis-configured" routers. I need
> >> to get some coffee and hopefully my brain will get in gear....
> >>
> >
> > I believe that was ECN. (A way to signal congestion without throwing away
> > packects.)
> >
> > There have also been problems with MTU discovery because people block ICMP
> > packets indiscrimiantly.
> >
> >
> That is starting to stir the synapse....
>
> The problem is related to what is in the "syn, ack" packet coming from
> 12.17.249.39. In the TCP Options it is sending "Window scale: 0
> (multiply by 1)". I think this is causing problems since later packets
> have the TCP data segment being only 92 bytes from 12.17.249.39.
>
> While the same info is in the "syn, ack" the RHELv4 system seems to deal
> with it OK and the TCP data segments from 12.17.249.39 are 1460 bytes.
>
> The "problem" can be fixed by....
>
> echo 0 > /proc/sys/net/ipv4/tcp_window_scaling
>
> Now, while this fixes the problem for this site it may cause performance
> issues with other sites. I don't know at this point and my head hurts.
> :-) It is hurting since my RHELv4 system has tcp_window_scaling set to
> 1 and it seems to deal with it.
>
> Note: Make sure you issue the "echo" above after some quite time. The
> connection must first be rst (reset) before the sides will renegotiate
> the TCP/IP settings. (That too caused me some grief.)
>
> I'll bugzilla it....
What became of the fix for this issue ?
The reason I ask is because I'm now running Fedora 14 and a) I still
have the problem and b) about every 3 days I have to edit the
tcp_window_scaling file and reset the scaling entry to zero. Something
over writes it back to true.
Thanks !
More information about the users
mailing list