Hi Rick,
On 29/09/14 19:28, Rick Jones wrote:
On 09/27/2014 01:59 AM, Manolis Sifalakis wrote:
A couple of remarks to try and contribute to this discussion, one
philosophical and one engineering.
The philosophical first. There is a presentation taking place in a room
and it is possibly at q&a phase. You have several ppl entering the room
at several time points. What if every person entering the room,
immediately throws aggressively questions without even having spent time
to listen and understand the context of presentation and discussion.
Imagine this happening again and again. Consider what are the chances
that a meaningful communication and discussion will take place... or
simply how would you like that as a person in the audience?
I believe I see where you are going, but the analogy is missing one key
piece. That is that before entering the room, each person must make a
request to enter (send a TCP SYN), which must be acknowledged by the
presenter (SYN|ACK). When the presenter gives each person permission to
enter, she also tells them how many questions they may ask, in the form
of the value for the (initial) TCP receive window in the SYN|ACK.
When the presenter knows she has a number of questions already
outstanding, and/or she knows that the questions will be arriving
through a small pipe, she can tell each person entering the room they
may ask only a small number of questions.
Fair enough. Particularly when it comes to the control the receiver side
can exert.
On second thought however..., the recv window is intended to set an
upper bound (cap) to an otherwise nominally well behaving sender, and
not to regulate the aggressiveness of the sender (init wnd and rate of
growth). The two are similar but not the same. I m not sure about the
effects of managing them in tandem, since setting the cap will not
necessarily resolve the effects of aggressive behaviour.
The more technical next. Doubling and tripling IW for short lived
sessions (each of which will attain --only-- exponential growth in its
anyway short lifetime) means that a large number of high-freq transient
components will be added to and removed from the signal that the AQM
sees (and tries to adapt to). Is that something desired ? Will it
improve or worsen the way an AQM adapts? Increasing IW is not only a
matter of initial value, it also affects the starting growth rate (since
the relation is not linear).
I have a sneaking suspicion "we" (the collective group of folks working
on "the internet" as it were) went through a similar debate over IW3
versus IW1. And ten years from now we will probably have a similar
debate over IW30 versus IW10 :)
possibly! .. and it is habitual for the older to be more conservative ;)
My understanding was the concern put forth was that a slow receiver
might be overwhelmed by communicating with a well-connected sender using
IW10. And someone suggested there should be a signal the slow receiver
could send to the well-connected sender that said "Don't do IW10." All
I am saying is that if we do indeed have a slow receiver, it already has
such a "signal" - the TCP receive window. Even if the well-connected
sender was using IW1000, it must still love, honor and respect the
classic TCP receive window.
I see your point.. and the effect you advocate.
I m just not convinced of the philosophy "Let em be aggressive, we can
always call police *after* should there be need". This can have
side-effects that "police" may not be effectively addressing (for
en-path queueing par example).
How well an AQM will like or adapt to a bunch of loud-mouth mice versus
a soft-spoken elephant I do not pretend to know. But while I have a
soft spot for some quixotic things - for example, the (IMO) mis-use of
impact and impacted :) and will complain about other things unlikely
ever to change (for example, the Linux TCP stack's seeming willingness
to go "Grow the receive window huge and let congestion control sort it
out." I really, really doubt the IW10 genie is ever going back into the
bottle. I'm also not really all that upset with IW10 in the first
place, but then I was one who thought that counting the ACK of the SYN
or SYN|ACK when growing the congestion window was fine.
rick jones
M.
_______________________________________________
Bloat mailing list
[email protected]
https://lists.bufferbloat.net/listinfo/bloat
