Re: [tcpm] New Version - draft-ietf-tcpm-newcwv-07.txt

["Scharf, Michael (Michael)" <[e-mail]>]

> So we've now had time to revisit the changes to Linux and the
> differences between the ID and the Linux code.
> 
> A patch was submitted and followed the ID. This was available up to
> Linux 3.15.
> 
> In Linux, there was an update submitted that conflicted with the
> new-cwv patch, so new-cwv is not now supported - we think this is what
> was recently discussed on the list. From what we understand, the
> conflicting update centred around fixed needs to support TSO. This was
> explained in the email from Yuchung on 22/09/2014 describes this work
> from Neal Cardwell.
> 
> To us, it seems this does not implement a window validation procedure,
> so this isn't the same as new-cwv. We'll highlight what we see are the
> key differences below, and others may comment - especially if we get
> this wrong.

Thanks! This analysis is very useful.

Here is a random thought: I wonder if one could describe new-cwv as a set of modular components/algorithms, instead of the current "all-in-one" algorithm. Each of these components could perhaps be implemented in more than one way, obviously resulting in different trade-offs.

Actually, I doubt that there has every been an agreement among TCP stacks how to exactly perform congestion window validation, and I don't know if even inside TCPM we will ever have such a common understanding.

Thus, as a remedy to the current situation, would it be possible to put both new-cwv and the Linux approach (and possible other algorithms) under a common umbrella? Such a new-cwv I-D could first describe the basic functions that have to be solved (i.e., a framework), plus some example algorithms (i.e., the well-known variants we are aware of).

Any thoughts?

Michael (as individual contributor)