[e2e] TCP Loss Differentiation

[David P. Reed]

Before going too far in this direction, one should note that unicast 
traffic on layer 2 transports commonly used in practice for Internet 
transport has negligible loss rates, even on wireless networks such as 
802.11.

The problem of differentiation arises when attempting to elide layer 2 
functionality and run "TCP/IP on bare PHY".  Otherwise "link loss rate" 
is a concept without much reality at layer 3.  We don't run TCP/IP on 
bare PHY layers.  We run it on layer 2 protocol, over PHY layers, which 
protocols always have high reliability today.  Some  multicast layer 3 
protocols run on unreliable layer 2 multicast protocols (such as 802.11 
multicast), but TCP/IP never uses multicast.

Layer 3 losses are nearly always the result of *only* 2 very different 
phenomena: 1) buffer overflow drops due to router/switch congestion 
queue management or 2) layer 2 breaks in connectivity.

Thinking about "link loss rates" is a nice academic math modeling 
exercise for a world that doesn't exist, but perhaps the practical 
modeling differentiation should focus on  these two phenomena, rather 
than focusing on "link loss rates".  The "connectivity break" case 
(which shows up in 802.11 when the NIC retransmits some number of times 
- 255?) doesn't have very good statistical models, certainly not the 
kind of models that can be baked into TCP's congestion/rate control 
algorithms.  And that model is not likely to be poisson, or any 
distribution easily characterized by a "rate parameter".


Detlef Bosau wrote:
> Hi.
>
> Some years ago, the issue of packet loss differentation in TCP was a 
> big issue. Does somebody happen to know the state of the art in this 
> area?
>
> I'm particularly interested in those cases were we do _not_ have a 
> reliable knowledge about the loss rate on a link. (So, particularly 
> the CETEN
> approach by Allman and Eddy cannot be easily applied.)
>
> Thanks.
>
> Detlef
>