Re: [Spud] [Stackevo-discuss] Fwd: Possible WG-forming follow-on to SPUD BoF

[Brian Trammell <[e-mail]>]

> On 23 May 2016, at 18:00, Joe Touch <[e-mail]> wrote:
> 
> 
> 
> On 5/23/2016 6:44 AM, Brian Trammell wrote:
>>> On 20 May 2016, at 18:03, Tom Herbert <[e-mail]> wrote:
>>> ...
>>> DSCP is properly network-layer. If you believe in a "path layer" (which I'm pretty sure I do), you can argue that ECN is the first path layer protocol.
>>> 
>>> I'm not sure what "path layer" is. Yes every packet takes some path,
>>> but the Internet is packet-switched not circuit-switched, so not all
>>> packets for a flow are required to always take the same path.
>>> Maintaining flow state in intermediate nodes is only best effort. In
>>> the presence of multi-homing and mobility intermediate devices that
>>> track flow state will inevitably have it wrong. Anyone who has ever
>>> tried to build a front-end L4 load balancer understands this problem
>>> all too well :-)
>> Your L4 load balancer is a path layer device.
> 
> The only notion the Internet has of a path is a flow label.
> 
> Transport IDs identify services and connections within an endpoint, and
> are not universally available or meaningful except at the endpoints.
> 
> All attempts by intermediate devices to rely on those identifiers makes
> assumptions about visibility and correlation to desired shared-path
> behavior that are baseless.

But those intermediate devices *do indeed* route those packets based an N-tuple, where N is generally 5. One can ignore this fact and be surprised by what the network does, or resign oneself to it and leverage it.

> This is the key reason why MP-TCP is very badly misnamed. It is
> multi-endpoint, not multipath, and always has been.

You're quite right, but the current routing architecture gives us no way to influence paths except by choosing endpoints. And even then, we can know very little about how the endpoint identifiers map to paths.

Cheers,

Brian