[tsvwg] L4S issue #16/17 questions from reading the session slides

[Sebastian Moeller <[e-mail]>]

Dear tsvw community,

with great interest I had a look at the slides Bob posted earlier. I have a couple of questions I would like to discuss:


Issue #16: Fall-back to Reno-Friendly on Classic ECN bottleneck

A) The slide ends with the observation:
 "	• detection unlikely to be perfect"
I now want to ask, which traffic class is supposed to pay the price for this detection misses? The principle of least surprise would indicate that the detector should err on false positive RFC3168 detections.

QUESTION: Does everybody agree that it should be safety first and the cost of mis-classifications are carried by L4S endpoints/flows?
(In that case the logic of the detector should be assume RFC3168 until there is sufficient evidence to rule that out with a certain level of confidence)

B) "Active detection technique
if last two reordered, likely L4S ● reduce suspicion, and continue"

Or a bonded link...

QUESTION: Is it permissible for the detector to be fooled by bonded link segments (especially in the light of L4Ss push for time based retransmission like RACK based on the existence of bonded links)?

C) The alpha bug:
"All prior testing had been done using DCTCP.
Upon fixing this bug, the results again match DCTCP"

QUESTION: Is it really a good strategy to assume tests performed with dctcp carry over to other transport/CC algorithms? Or does this indicate that the published results so far using dctcp should at least be confirmed with TCP Prague?


D) "Issue #17: Conclusion
The main result of concern was due to a bug in initializing the value of TCP Prague alpha, which has been fixed and demonstrated to resolve the latency impact that was spanning multiple seconds"

COMMENT: In the light of https://www.heistp.net/downloads/sce-l4s-bakeoff/bakeoff-2019-11-11T090559-r2/l4s-s6-2/batch-l4s-s6-2-prague-vs-cubic-50Mbit-80ms_fixed.png showing that L4S still mishandles CE markings in that condition (albeit with the fall-out restricted to itself) it seems that the alpha fix alone might not be sufficient (which basically makes TCP Pragues response to the first CE independent of the CE interpretation scheme)


E) "If the CoDel queue is upgraded to perform Immediate AQM on L4S flows, the latency spike can be largely avoided."

QUESTION: True, but is proposing to change millions of deployed hosts really a viable strategy for coexistence/backward compatibility? 



Best Regards
	Sebastian