[e2e] Can feedback be generated more fast in ECN?

[Eric A. Hall]

> > Generating new packets packets needs vastly more computing than
> > marking passing packets, unless your computers are of the sort that are appear
> > only in cloud charts.  I'm completely flabbergasted by the repeated claims
> > in this thread that generating new packets is other than between expensive
> > and nearly impossible (e.g. in unidirectional pipes).
> >
> > Sheesh!--why have so many boxes for so many years mishandled other ICMP
> > messages, such as Echo Requests?  (e.g. by stupidly just swapping source
> > and destination IP addresses instead of putting the responding system's
> > address in the source field).  Do you guys really think that Cisco is
> > merely being mean and nasty by needing 50-100 milliseconds to generate an
> > ICMP Echo-Responses in routers that need microseconds to pass packets in
> > the forward direction?
> >
> > How do you fit the fact that host and router performance is much
> > better measured in packets/second than bits/second with the notion
> > that generating ICMP packets is as cheap as marking packets more
> > simply than the decrementing of the TTL?
> >
> > Even worse about the statements in support of source quench is the
> > fact that there is real life experience with them.
> 
> The single most difficult to do in almost all modern networking
> equipment is to generate packets in fast path. Invariably the handling
> of SQ has to be done in the slow path. Most of the routers have slow
> paths which is *really slow*. Any bandwidth consumed in the slow path by
> SQ is bandwidth taken away from processing of routing updates, router
> alerts, option processing and a whole bunch of other traffic.
> 
> Vinay

-- 
Eric A. Hall                                        http://www.ehsco.com/
Internet Core Protocols          http://www.oreilly.com/catalog/coreprot/