lkml.org 
[lkml]   [1998]   [Sep]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: TCP Stalls.
Date
From
Alan Cox wrote:
> > If I corrupt loopback so it fails to transmit a few packets, I can
> > emulate the same kind of TCP stalls that have frequented the kernel.

> Your model is then flawed because the back off doesnt induce lower error
> rates through reduced congestion. You model a continuous 25% loss link
> layer - and that is something that unaided is beyond the capabilities
> of an IP network.

Oh Alan, but his model IS correct in that respect, considering the
modern internet.

Typically, I have a "clean"(*) 5 hop route to a link that's
permanently congested. There are thousands of people behind that link,
and they are ALL starting new connections (with TCP stacks that don't
know the congestion stats for that link yet) in a rate that will keep
the link congested.

1. rosie 0% 108 108 1 1 6
2. ci-gv1.casema.net 0% 108 108 45 112 298
3. ci-gv1.casema.net 0% 108 108 47 117 277
4. dh-inet.casema.net 0% 108 108 45 127 319
5. wdc.core.ixe.net 1% 106 108 237 998 2439
6. mae-east-gw1.cableol.net 6% 101 108 207 634 1749
7. linx-gw1-lon-ATM1-0.router.cableol.n 6% 100 106 318 698 1711
8. betelgeuse-ATM5-0.router.cableol.net 3% 104 107 335 724 1820
[6 more hops]

This is from here to www.ukuu.org.uk As you can see, hop 4-5 and hop
5-6 are the ones that suffer from packetloss. Actually, the situation
at this moment is pretty good: only 6% packetloss.

In these situations TCP specs say that you need to back off. In
practise that is not good for performance. Usually performance
drops to the "100 bytes per second" range.

I've written a program that will send UDP packets at a fixed rate.
My link then does about 10k per second with 10% packetloss. Overall
I can get 9k per second through.

A nastier hack (violates RFCs, but is less demanding on the internet
links of my provider) is to have TCP send the latest ACK every second
no matter what. This will trigger a fast retransmit after about 3
seconds instead of after a longer and longer delay.

Roger.


(*) < 2% packetloss

--
| The secret of success is sincerity. Once you can | R.E.Wolff@BitWizard.nl
| fake that, you've got it made. -- Jean Giraudoux | T: +31-15-2137555
-We write Linux device drivers for any device you may have! Call for a quote-

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.altern.org/andrebalsa/doc/lkml-faq.html

\
 
 \ /
  Last update: 2005-03-22 13:44    [W:0.077 / U:0.940 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site