[e2e] Is a non-TCP solution dead?

[Michael Welzl]

Dear Alex,

I've been following this for a while and I finally decided
to add my 2 cents:


> 1) We do not know the network layer underlying any flow with certainty.  Thus,

Right!
So the Internet is about connecting everything and everyone, no
matter how heterogeneous, isn't it? IP over ATM VCs as well as
avian carriers.


> for instance, your ATM example is quite possibly relevant to many paths that
> one might use.  We simply don't know when rate controls are being employed
> below TCP without having access to, and analysis of, the paths.  So, we lose
> performance by not taking advantage of good network layer congestion
> management, because TCP assumes error loss is congestion loss. This is
> particularly true in private networks.

Right. So here's how this problem is being dealt with:

Use mechanisms like source quench -> usually doesn't scale
maybe explicit corruption notification ... well, this is being
discussed occasionally ... see the trigtran effort, for instance -
I think the most difficult part is to make such a mechanism scale.

Another possibility: do congestion control somewhere along
the path - well, there are PEPs and lots of  mechanisms that
split a path to deal with the effects of wireless links ...
it of course breaks the control loop and doesn't work as well as
TCP would if we had all this functionality in TCP.

On the other hand, TCP doesn't know if it's faced with a
[wireless, ...] link. So inform it - well, we're back at
the source quench scalability problem again ...

So, TCP is a general-purpose solution which doesn't know
anything about lower layers. That's the assumption that
we will just have to live with.


> The bottom line is that some of TCP's flow-control complexity is unnecessary
> in end-end transport management, but was added in years ago to try to protect
> network-layer nodes from disaster.  Since we now have more unreliable parallel
> flows and more unreliable paths, as in wireless, to contend with, TCP is an
> inadequate solution to a growing network-layer problem.

Quite a statement. Which general-purpose solution do you propose?

Cheers,
Michael