[e2e] Is a non-TCP solution dead?

[Cannara]

Nicolas, just a correction or two (and one for Hans, to save emails)...Alex

Nicolas Christin wrote:
> 
[clip]
> I am afraid numbers of backbone link utilization will not lead any
> insight w.r.t. TCP performance, because of overprovisioning. On the
> other hand, I'd be *extremely* interested in seeing *proof* that non-TCP
> (or "fake" TCP) traffic is increasingly dominant, as Alex was initially
> implying, before seemingly backing off in his last post.
...I wasn't arguing the current weighting of traffic, only pointing out the
relationship effects, should the weighting go one way or the other, as it
appears to some it is.  As I observed, if pkts are dropped on a TCP session,
the slowdown is excessive for the drop rate.  The same drop rate, usually from
the ISP edge, affects non-TCP flows far less.  They, in turn, can affect TCP
more.  TCP flows affect each other as well.  In any case, the issue is with
TCP and its assumptions.  As an aside, consider that what we might want to
research is how to better handle a network whose peering points often drop 25%
of incoming pkts -- e.g., see the SLAC reports by Les Cottrell et al.

For Hans, dumping Slow Start doesn't mean not responding to losses or
congestion.  If all a TCP had excised were Slow Start (nssTCP), then it would
still have a retrans timer that, say, doubled for every unAcked retrans.  So,
whacking the peering points with all this initial traffic would indeed incur
heavy loss, as current traffic now does at busy times, and escalating retrans
delays would reduce load (just as they do/did for transports other than TCP).
If we were actually in an ECN world, then TCP could be made to do even better,
but then, so could IP and thus all traffic -- the real point here.

The issue of overprovisioning and backbone traffic % is economic and historic
-- ISPs and other Internet-access sites (Covad, etc.) have to market last
mile service at highest possible data rate, but have limited telco-office
pipes to source/sink onward.  The common business model for DSL was 20:1
oversubscription, for this reason.  Yes, 20:1.  This is why DSL users often
complain and why pkts at the edge are dropped (thus lightening the load within
the core).  See, for instance: "The trouble with DSL" a year or so back.  The
core is, indeed, overprovisioned, which is fine.

[clip]
> The other problem I see with this thread is that the argument seems to
> change every time a new post is made. From what I read, "TCP performance
> sucks" became "Traffic is becoming non-TCP/fake TCP" before rolling back
> to "TCP performance sucks". OK, fair enough, Alex, but please back this
> argument with *data*, otherwise, it just reads like a disorganized rant.
..As I said, my aim was always open examination of better network transport
through better network management, at the network layer, which was not
considered in the design of the Internet protocols.  I really don't care for
"fake" TCP or whatever -- having such traffic work successfully simply
indicates another odd, but allowed, behavior descendent from the lack of
security designed into the Internet.  My argument has always just been the
above.

Alex