[tcpm] ECN and ECN++ traversal experiments: draft-ietf-tcpm-generalized-ecn-02

[Bob Briscoe <[e-mail]>]

TCPM folks,

We've updated this draft, mainly to add a ref to the paper below, which 
outlines a large measurement study on traversal of ECN++ and ECN.
It is the first major study of ECN traversal over both mobile and fixed 
networks.

It's "to appear" in IEEE Comms Mag, but here's a link to the accepted draft:
"Measuring ECN++: Good News for ++, Bad News for ECN over Mobile 
<http://www.it.uc3m.es/amandala/ecn++/ecn_commag_2018.html>"

It's fairly short, but here's an even shorter summary, solely of the 
results.

*Terminology: **
** outgoing -> from client to server; otherwise incoming.
* Classic TCP/ECN:     All TCP control packets and retransmissions are 
not ECN-capable
* TCP/ECN+      TCP SYN/ACK is ECN-capable, but other TCP control 
packets and retransmissions are not
* TCP/ECN++     all TCP packets are ECN-capable

*Highlights:**
*Bad news: 7.5 out of 11 mobile operators tested bleach ECN in outgoing 
packets at or before the first IP hop. {Note 1}
Good news: ECN fields do not cause packet drop.
Good news: ECN++ support is as good as ECN support.
Good news: 61% of Alexa top 500k Web servers support ECN.
Good news: All the ECN-capable top-500k Alexa servers we were able to 
test have the same response to ECE as to 3 DupACKs.
Good news: At least, 51% of the Alexa top-500k servers support IW of 10 
segments.
Bad news: 0.4% of the Alexa top-500k servers (1,745 servers) use IW 
larger than 10.
           Of these, 1,121 servers deliver the whole file in the IW 
(largest: 585 packets * 100 B = 58,500 B)
Bad news: Apparent ECN+ servers respond oddly (IW + 1) to congestion 
marking on the SYN/ACK: 3.51% of Alexa 500k


____________________
{Note 1}: Supplementary details related to the first bullet:
* Limited outgoing tests over 7 more mobile operators, find 3 more that 
bleach all ECN at or before the first IP hop.
* The "0.5 of an operator" bleaches ECN only on port 80 (non-ECN Web 
proxy).
* 2 operators use ECN-capable proxies.
* Over the 3 mobile networks that do not bleach ECN,
     0.53% of upstream paths to Alexa-500k suffer ECN bleaching >5 hops 
from the client
* Over fixed access (PlanetLab)
   - no upstream ECN bleaching close to client
   - 0.23% of upstream paths to Alexa-500k suffer ECN bleaching >5 hops 
from the client
* No downstream ECN bleaching measured over fixed or mobile,
   - except the non-ECN proxy,
   - however, limited paths tested (limited to servers we controlled)


Bob and co-authors