On Fri, Aug 12, 2016 at 10:59 AM, Joe Touch <[email protected]> wrote:
>
>
> On 8/12/2016 10:54 AM, Tom Herbert wrote:
>> On Fri, Aug 12, 2016 at 10:21 AM, Joe Touch <[email protected]> wrote:
>>>
>>> On 8/11/2016 4:59 PM, Jesse Gross wrote:
>>>
>>> The most common example given in this area is the use of IP options. At the
>>> time that routing was moving to hardware implementations, options were not
>>> widely used and so were not implemented. However, imagine that options were
>>> in common use – do you think that router vendors would have decided that IP
>>> was too difficult to implement and abandoned that market? Or would we now be
>>> accepting that this is a common element of protocols?
>>>
>>>
>>> A good example from history here are TCP options. In the beginning, they
>>> were considered only a complexity overhead.
>>>
>>> Another example is IPsec.
>>>
>>> Both, are examples of TLV options (TCP within the TCP header; IPsec as a
>>> 'next layer' protocol).
>>>
>>> Both are widely supported in hardware.
>>>
>>> That's not to suggest that we should focus on TLV solutions, but it goes to
>>> prove that even TLV-format options do not prevent widespread hardware
>>> support.
>>>
>> Joe,
>>
>> I don't know sure that parsing of TCP options is really all that
>> widespread,
> In hardware at the endsystems, it is.
>
>> but I do know that the inability of middleboxes to
>> properly implement support for both IP options and IPv6 extension
>> headers have pretty much made those non-starters for widespread
>> deployment.
>
> IPsec is a counterexample.
>
> For many years, the network device community pushed back on options
> because it undercuts their profit model (build for the 80% case and toss
> the rest to software slow-path).
>
> The tide there is turning; even the OPS groups are starting to admit
> that we can't simply deprecate IP options, but we can do things to
> encourage their more widespread support (e.g., limit the chain, etc.).
>
I hope you're right and extension headers do become deployable someday...

> Middleboxes ignore or misimplement options by design; they''re never
> useful as a constraint for protocol design.

Wouldn't placing a limit on the header length chain as you alluded to
above be an example of such a constraint on protocol design?

Tom

>
> Joe

_______________________________________________
nvo3 mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/nvo3

Reply via email to