Messages in this thread | | | Date | Fri, 11 Sep 1998 18:36:43 -0400 | From | "Theodore Y. Ts'o" <> | Subject | Re: my broken TCP is faster on broken networks |
| |
From: smurf@noris.de (Matthias Urlichs) Date: 11 Sep 1998 23:24:59 +0200
> Congestion is not a temporary problem that can be fixed in such > a simple manner. Even if every client does the backoff (hah!) > there will be plenty of packets to which it does not apply. > ... which is why the backbone routers now start doing the backoff algorithm themselves: the more packets you send them without regard for proper backoff algorithms, the more likely it is that _your_ packet will get tossed.
Indeed, Van Jaconbson presented an insightful talk on this very subject at an TCP implementors meeting at the IETF about 9-12 months ago. His thesis was that now that the Internet is growing, our congestion avoidance schemes which depended on everybody being polite was in danger of falling apart. All aside from the Alan's point about humans hitting the stop key on their browser and then resending their link, there is a "tragedy of the commons" effect where implementors who break exponential backoff win at the expense of people who obey the RFC's --- but if everyone did this, the net would melt down.
One of his suggestions was to have the core routers enforce backoff, so that the incentives for cheating would go away.
I don't normally follow the TCP implementor's working group (*), so I don't know if they actually followed through on his suggestion; Matthais's note seems to indicate that they did.
- Ted
(*) I only went because I heard that Van Jacobson was speaking, and he's a truly wonderful speaker that can make the most subtle points about TCP easy to understand. If you ever have an opportunity to hear Van Jacobson speak, I'd strongly advise you to go and listen to him. You won't regret it.
- 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.tux.org/lkml/faq.html
| |