[e2e] TCP ex Machina
Jon.Crowcroft at cl.cam.ac.uk
Sat Sep 21 00:43:52 PDT 2013
In missive <523CCCE3.3050209 at detlef-bosau.de>, Detlef Bosau typed:
>>And I really wonder, why we did not think that way the last 20 years.
we did - longer - since DECBit, ECN, re-ecn, congestion exposure -
lots of people did.
>>End points may report end point conditions. Network conditions must be
>>provided by the network. Not by endpoints, be it an educated guess, some
>>sophisticated conjecture or simply throwing dice.
however, notwithstanding the failure to deploy lots of smart
thinking to get explict congestion feedback out of the nodes in the
net, end system inference is not an _educated guess - its a
perfectly good statistical technique, and there have been lots of
steps refining the inference based techniques to improve end system
CC in the absence of network exposure of explict congestion
information - obviously they arent perfect, but neither is ECN -
without instantaneous action-at-a-distance, ANY distributed
technique has to cope with latency and noise - so by the time a
source received explicit information about network conditions, it
is out of date already, and said explicit information may not be
100% reliably delivered - so ECN and conext are also just
refininements, in the way fancier controllers and filters are just
refinements....just the explicit feedback is potentially a
slightly bigger refinement
note that congested nodes in the net can't predict what a bunch of
end node/traffic sources are _about_ to do either, so you still
need all the feedback control loop paraphanalia all over the
place....its just good engineering...
More information about the end2end-interest