Re: Is Fragmentation at IP layer even needed ?

[Joe Touch <[e-mail]>]

On 2/12/2016 11:10 AM, Warren Kumari wrote:
> 
> 
> On Fri, Feb 12, 2016 at 1:27 PM Joe Touch <[e-mail]
> <mailto:[e-mail]>> wrote:
...
>     Routers shouldn't reassemble, but then routers aren't supposed to look
>     beyond L3. You cannot have it both ways.
> 
> You keep saying that.... and then a bunch of operators say "Yeah, but I
> have an actual network to run, and I need to look beyond L3 because my
> customers want me to mitigate their DoS, I want to filter on L4 before
> handing data to internal services, and I use ECMP and need L4 because I
> cannot rely on flow labels".

Operators who don't want to pay for the devices that will actually do
the work involved to support the filtering model they want to employ.

This is a lot like selling empty boxes of cereal - because I need to
sell boxes of cereal, and the cost of the food is getting in the way.

>     Once you inspect L4, you *are* acting as a host.
> 
> So, this entire thread (which has reminded me why I stopped
> participating in v6ops) is just a terminology issue? ;-)

It's about what level of work you should be expected to expend to
produce a desired behavior. Ultimately, it's about the conceptual
Internet architecture.

...
>     As Mark pointed out, you don't need to strictly reassemble (i.e., to
>     emit a corresponding reassembled packet). You just need to reassemble
>     the information.
> 
> Which requires keeping state, yes? This is not realistic in modern large
> network devices. 

Translation: finding a place to keep all that cereal isn't realistic
when what I really want to do is make money.

Sorry - ante up. If you want to have the benefit of acting like a host,
you need to do the work of a host.

And note that some commercial devices do already support this sort of
state, so it's not theoretical.

> Saying "vendors should jolly well do a better job and redesign their
> gear so that it can, and operators should simply pay whatever the cost
> is... oh, and redesign their networks for flow consistency, because
> *that's what the specifications require*" is likely to continue having
> people say "Yeah, sure, whatever. But I've got a real network to run...."

There are real profits to be made. All that cereal is getting in
the way. It's a lot more profitable to sell empty boxes.

The real issue here is the lack of compliance testing and certification.
If we had that, vendors wouldn't be able to sell boxes that didn't
comply with requirements and operators wouldn't be able to make promises
they can't keep.

...
>     Again, the model leads you to the correct conclusions.
>
> ... and yet we see lots of evidence that fragments (and EH) have issues
> in real world testing.

What we see is operator and vendor greed being incompatible with
Internet architecture. We need to push back on *that*, not water down
our definition of the Internet to support their current business model.

Joe