[e2e] Bufferbloat

[Detlef Bosau]

Jon,


basically, the term Bufferbloat seems to me more like an agreement than
a hard fact.

In any queueing system we observe the sojourn time = service time +
waiting time.
And when I correctly remember DPR's clarification some months ago, the
term Bufferbloat means nothing else than that a jobs queueing time
exceeds the jobs service time.

Unfortunately, from an end to end point perspective, neither of these
can be assessed. We can assess the sojourn time, we can assess the
throughput. We cannot (at least not in the general case) assess a jobs
service time or a jobs waiting time.

Hence, RJ's perfomance metric is a derived term which maps assessable
metrics (sojourn time, throughput) to some number suitable as input for
some control mechanism. (This is a generic problem in control theory, VJ
refers to Luenberger in his congavoid paper, refer to the textbooks of
Luenberger or Kalman here. Keywords: Luenberger Observer, Kalman Filter.
The problem is how we assess variables which cannot be directly observed.)

When we coin a term for the situation "waiting time exceeds service
time", this is an agreement. Nothing like an "universal constant".

However, any approach to avoid buffer bloat from an end to end
perspective is restricted to one control variable (the congestion
window) and to derived assessment variables (see above) as well.

This is a direct consequence of building a feed back control loop here.

Could it be worthwhile to discuss an open loop, feed forward approach as
an alternative?

-- 
------------------------------------------------------------------
Detlef Bosau
Galileistraße 30   
70565 Stuttgart                            Tel.:   +49 711 5208031
                                           mobile: +49 172 6819937
                                           skype:     detlef.bosau
                                           ICQ:          566129673
detlef.bosau at web.de                     http://www.detlef-bosau.de