Re: [tcpm] New Version Notification for draft-touch-tcpm-sno-00.txt

[Joe Touch <[e-mail]>]

Hi, Rui,

On 9/9/2013 11:57 AM, Rui Paulo wrote:
> On 4 Sep 2013, at 17:34, Joe Touch <[e-mail]> wrote:
>
>>
>> Hi, all,
>>
>> Regarding the recent discussion on TCPMUX, we've revived some of
>> our earlier work on portnames, focusing on just the ability to
>> decouple the service identifier from the SYN destination port.
>>
>> We have both Linux and FreeBSD implementations underway. A revised
>> draft has just been submitted with the details, below.
>>
>> I'm hoping to publish this as an independent submission, but if
>> there's interest in running it through TCPM please let me know.
>
> From a quick read, I only see one use case mentioned: proxy servers.
> Are there others?

It could come anytime a single client has a large number of connections 
to a single server. The proxy-to-proxy case is the expected one, but 
that could happen anytime the per-connection rate exceeds a fairly small 
number (around 550/second).

> Changing the TCP stacks of these endpoints just to accommodate a
> bigger number of connections seems like the wrong approach when you
> can rework the network architecture and add more proxy servers.

This solves the issue regardless of how proxies are deployed. It may be 
the "wrong approach" to solve the problem, but arguably overloading the 
SYN destination port with two meanings was the wrong approach in the 
first place.

This is attempt to show how to solve the problem in a clean way. You can 
argue that it's not clean enough, but I'd then ask you to architect a 
solution and document it as an alternative.

FWIW, I doubt this will get a lot of use; it's currently targeted as 
experimental.

Joe