IETF
iccrg
iccrg@jabber.ietf.org
Monday, March 23, 2015< ^ >
jishac has set the subject to: Internet Congestion Control Research Group
Room Configuration
Room Occupants

GMT+0
[13:56:31] Meetecho joins the room
[13:56:56] John Leslie joins the room
[14:01:58] scribeJohn joins the room
[14:06:15] Kiran YEDUGUNDLA joins the room
[14:06:43] Brian Trammell joins the room
[14:07:15] <scribeJohn> I assume Brian is on-site...
[14:07:27] <Brian Trammell> yep
[14:07:35] <scribeJohn> in-room?
[14:07:38] <Brian Trammell> yep
[14:07:53] <scribeJohn> Gorry was supposed to be trying to jabber-scribe
[14:08:22] Greg White joins the room
[14:09:16] Gorry Fairhurst joins the room
[14:09:31] <scribeJohn> hi, Gorry
[14:10:00] <Gorry Fairhurst> Hi - didn't have a jabber client on my new laptop.
[14:10:16] <Gorry Fairhurst> Are any of the people in the jabber room remote participants?
[14:10:37] <scribeJohn> (Though I'm using MeetEcho, I'm using the expanded video, so I'm not sure what slide is displayed. :^(
[14:11:06] <Gorry Fairhurst> We're at slide 6 of CoCoa
[14:11:11] <Gorry Fairhurst> Slide 7
[14:11:17] <scribeJohn> thx
[14:11:18] <Gorry Fairhurst> Slide
[14:12:21] <Gorry Fairhurst> Network diagram
[14:16:42] Brian Trammell leaves the room
[14:16:54] Brian Trammell joins the room
[14:17:29] Gorry Fairhurst leaves the room
[14:17:56] Gorry Fairhurst joins the room
[14:18:01] <Gorry Fairhurst> Slide 11
[14:19:40] <Gorry Fairhurst> Slide 12
[14:20:34] <Gorry Fairhurst> Slide 13
[14:21:45] Todd Tolbert joins the room
[14:22:17] <Gorry Fairhurst> next slide (fairness)
[14:22:47] <scribeJohn> yeah... the slide numbers are hidden sometimes...
[14:25:11] <Gorry Fairhurst> next slide (settling)
[14:25:17] Todd Tolbert leaves the room
[14:27:42] <Gorry Fairhurst> Slide 17
[14:28:35] <Gorry Fairhurst> Sldie 18
[14:30:35] <Gorry Fairhurst> Slide 20
[14:32:08] Kiran YEDUGUNDLA leaves the room
[14:35:53] <Gorry Fairhurst> Questions?
[14:37:45] <Gorry Fairhurst> Richard S at mic:
[14:37:47] Aaron Falk joins the room
[14:38:18] <Gorry Fairhurst> Thanks - next speaker: Koen
[14:40:21] <scribeJohn> Do I gather there are no slides online for this presentation?
[14:42:29] <Greg White> Is it possible to turn up the gain on the speaker's mic?
[14:47:19] Kiran YEDUGUNDLA joins the room
[14:50:47] <scribeJohn> (I can't make sense of the "throughput" slide....)
[14:58:17] Gorry Fairhurst leaves the room
[14:58:42] Gorry Fairhurst joins the room
[14:58:53] <Gorry Fairhurst> Slide 22:
[15:00:42] <Gorry Fairhurst> Slide 24:
[15:02:15] <Gorry Fairhurst> Slide 26:
[15:02:20] nagesh javali joins the room
[15:03:13] <Gorry Fairhurst> Slide 27:Q
[15:04:05] <Gorry Fairhurst> Slide 2: Throughput
[15:04:42] <Gorry Fairhurst> Slide 29:
[15:06:09] Aaron Falk leaves the room
[15:06:22] <Gorry Fairhurst> Slide 30:
[15:06:44] Brian Trammell leaves the room
[15:07:26] <Gorry Fairhurst> Slide 31:
[15:09:28] nagesh javali leaves the room
[15:10:32] <Gorry Fairhurst> Questions from Mic: Janna I.
[15:11:28] <Gorry Fairhurst> "Where did  come from? Did you do experiments with mixed RTT flows?"
[15:12:39] <Gorry Fairhurst> "1.22/2"?
[15:13:11] <Gorry Fairhurst> We assume a base RTT of 8ms, this is a FIXED time in the system we describe
[15:14:26] <John Leslie> mic: did you observe no differences between RENO and CUBIC?
[15:14:33] <Gorry Fairhurst> Slide 33 (extra)
[15:15:03] <Gorry Fairhurst> Question from Mic: David Black
[15:15:47] <Gorry Fairhurst> "The comes from fixed RTT assumptions - is there any way to set this on the fly to make sure it matches the network conditions?"
[15:15:57] NAGESH JAVALI joins the room
[15:17:07] <Gorry Fairhurst> Slide 29: "What happens in the ECN Class box"
[15:17:47] <Gorry Fairhurst> Mic: Stuart Cheshire
[15:18:21] <Gorry Fairhurst> "there are 3 variables: this is hard to compare, it depends on multiple changing variables"
[15:19:14] <Gorry Fairhurst> There is an opportunity to make a bigger change to ECN
[15:19:37] <Gorry Fairhurst> Stuart: All packets in the network should start to use ECN
[15:20:52] <Gorry Fairhurst> Stuart: I get the sense that the sharp cliff is an ideal, and my observation is that in a very heterogenous network, the packets will have bursty traffic patterns that may give short term variability in rate.
[15:21:39] <Gorry Fairhurst> Koen: Agree.
[15:21:43] <Gorry Fairhurst> Mic: Bob Briscoe
[15:22:58] <Gorry Fairhurst> Bob: ECN can identity more than ECN capability and I did a talk before that tries to re-purpose ECN.
[15:23:30] <Gorry Fairhurst> Bob: I think this needs to be much better than traditional ECN if we make this change.
[15:24:24] <Gorry Fairhurst> Bob: The good Q; bad Q argument is not valid for ECN - because this is a signal, that the endpoint can apply, V-J arguments apply only to drop.
[15:24:25] <Gorry Fairhurst> Mic Caitlin
[15:24:42] <Gorry Fairhurst> Caitlin: what happens if the clients move to the DC for example?
[15:25:21] <Gorry Fairhurst> Mic: Andre McG:
[15:25:40] <Gorry Fairhurst> … I agree with a lot that was said.
[15:25:56] Randell Jesup joins the room
[15:26:52] Randell Jesup leaves the room
[15:27:17] Randell Jesup joins the room
[15:27:18] <Gorry Fairhurst> … I think there are open questions. You must deal with the fact that a RTT can change - Cache nodes can be drained, and this then  causes a change you need to do this.
[15:27:34] <Gorry Fairhurst> Next Speaker ————————————————
[15:28:06] <Gorry Fairhurst> Next Speaker: Brian Trammell
[15:28:31] <Randell Jesup> barely audible... mic too far?
[15:28:44] <Gorry Fairhurst> Still a problem?
[15:28:49] <Randell Jesup> yes
[15:29:17] <Randell Jesup> better, thanks
[15:29:23] <Randell Jesup> low bu tlistenable
[15:29:34] <Gorry Fairhurst> Slide 2
[15:31:03] <Gorry Fairhurst> Slide 3:In the meantime… linux does passive mode ECN
[15:31:59] NAGESH JAVALI leaves the room
[15:32:12] <Gorry Fairhurst> Slide 4: Connectivity
[15:33:55] <Gorry Fairhurst> Slide 5: Endpoint
[15:34:52] <Gorry Fairhurst> Mic: Stuart Chesrire:
[15:35:10] <Gorry Fairhurst> Brian: This is server endpoint dependent
[15:35:43] <Gorry Fairhurst> Slide 6: Path
[15:36:03] <Gorry Fairhurst> Remote people: Audio level OK?
[15:36:43] <scribeJohn> fine for me
[15:37:05] <Gorry Fairhurst> Matt: we need traceroute that does ECN markings
[15:37:13] <Gorry Fairhurst> Brian: we have tracebox.
[15:38:48] <Gorry Fairhurst> Slide 8: results
[15:40:17] <Gorry Fairhurst> Slide 7: Results (sorry)
[15:40:25] <Gorry Fairhurst> Sldie 8: OS and Rank
[15:40:40] David Millman joins the room
[15:41:25] <Gorry Fairhurst> Mic; Jana:
[15:41:49] <Gorry Fairhurst> Brian, in this part only connectivity - CE later
[15:42:28] <Gorry Fairhurst> Brian: TTL plot shows "remaining" TTL, which may help identify platform OS
[15:43:15] <Gorry Fairhurst> Slide 9: Negoitiation
[15:44:27] <Gorry Fairhurst> Slide 10: Signaling
[15:45:32] <Randell Jesup> good.  even better than after earlier issue
[15:45:49] <Randell Jesup> (that's ref to audio)
[15:46:29] <Gorry Fairhurst> Mic: Matt:
[15:46:52] <Gorry Fairhurst> Matt: Are you aware that there are products that use these bits for things?
[15:47:27] <Gorry Fairhurst> Mic: Tim Shepherd: Comment on just 2 CE bits
[15:48:20] <Gorry Fairhurst> Brian: We didn't want to do a congestion-causing flow
[15:48:38] <Gorry Fairhurst> Return to slide 7:
[15:49:49] <Gorry Fairhurst> Mic Stuart Cheshire:
[15:51:11] <Gorry Fairhurst> Stuart: I've been running ECN for a while, look at wikipaedia and you can see how many people will talk ECN with you. I'm seeing non-zero stats on SWR received.
[15:51:24] <Gorry Fairhurst> Slides 11: Conclusions
[15:54:48] <Gorry Fairhurst> Questions?
[15:55:12] <Gorry Fairhurst> Mic
[15:55:39] <Gorry Fairhurst> Mic: Richard S: There are also "bleaching" issues?
[15:55:54] <Gorry Fairhurst> Brian :  This is a separate issue
[15:56:10] <Gorry Fairhurst> Mic Trevor Chandler:
[15:56:51] <Gorry Fairhurst> Trevor Chandler: Slide 7" 9.36% of IPv6 web servers are broken for Ipv6
[15:57:52] <Gorry Fairhurst> Brian: These are often broken IPv6 DNS, possibly no re-routing, and they seem to reply only on fallback to IPv4 Happy Eyeballs.
[15:59:30] <Gorry Fairhurst> Next talk ———————————————————————————
[16:00:20] <Gorry Fairhurst> Higher Throughput for TCP
[16:00:29] <Gorry Fairhurst> Slide 2: Motivation
[16:01:29] <Gorry Fairhurst> Sldie 3:
[16:03:33] <Gorry Fairhurst> Slide 4:
[16:04:34] <Gorry Fairhurst> Slide 5:
[16:05:06] <Gorry Fairhurst> Slide 6:
[16:07:14] <Gorry Fairhurst> Slide 8:
[16:08:58] <Gorry Fairhurst> Slide 9: Performance with only this traffic
[16:09:16] <Gorry Fairhurst> (no loss)
[16:09:37] <Gorry Fairhurst> Questions?
[16:10:22] <Gorry Fairhurst> Mic: Marie-Jose M.
[16:10:43] <Gorry Fairhurst> Why is this working, how does this not lead tp CC?
[16:10:46] <John Leslie> someone might ask if there's tracking of the time to recognize the loss...
[16:11:47] <Gorry Fairhurst> Queue at Mic: Marie-Jose; Bob Briscoe; Stuart Cheshire…
[16:12:55] <Gorry Fairhurst> MJM: I think this is not a valid change to TCP for deployment in the Internet. You need to look for other methods.
[16:13:11] <Gorry Fairhurst> Bob: This breaks other flows before they can respond.
[16:14:32] <Gorry Fairhurst> Stuart: Thank you for coming to the IETF and presenting this. Within a data centre this may be safe (simply sets a large IW) - this works until it breaks and then breaks the path for all other flows.
[16:15:02] <Gorry Fairhurst> Mic: Return to slide 23
[16:15:13] <Gorry Fairhurst> (slide 3)
[16:16:32] <Gorry Fairhurst> Mic Georgias: 4K video and UHD video will be deployed and this is important as a use-case.
[16:17:20] <Gorry Fairhurst> Targets of 50 Mbps are interesting to explore using TCP.
[16:17:47] <Gorry Fairhurst> Mirja: Slide 4: The red text is not true, this is why we have CC.
[16:18:05] <Gorry Fairhurst> Next presentation ————————————————
[16:18:33] <Gorry Fairhurst> Jinzhu Wang on network coding
[16:19:31] Kiran YEDUGUNDLA leaves the room
[16:21:33] <Gorry Fairhurst> Gorry: So there is a packet loss and you so no TCP RTO, but do you do a congestion response if there is a loss?
[16:21:35] <Gorry Fairhurst> <No>
[16:21:52] <Gorry Fairhurst> Gorry: So if you don't do a congestion resposne, this is not TCP.
[16:21:58] <Gorry Fairhurst> sldie:
[16:22:09] <Gorry Fairhurst> Slide: Coded TCP architecture
[16:24:22] <Gorry Fairhurst> Mic: Brian: Which bit do you want in the TCP Header?
[16:24:33] <Gorry Fairhurst> <A reserved field, or a TCP option>
[16:25:07] <Gorry Fairhurst> Mic: Bob: There are no bits that allow this
[16:27:20] <Gorry Fairhurst> Mic: Stuart Cheshire: This may be appropriate at a lower (wireless) layer, we can't just ignore the possibility though that the wifi loss is not congestion - if wifi is not reliable, this could be fixed in the wifi network - doing this over an internet path is not the correct way to do this.
[16:28:31] <Gorry Fairhurst> This is not a reasonable transport method
[16:30:18] <Gorry Fairhurst> Mic: Andrew McG: Only 1% loss can be common or higher, sometimes 45%. This could be used end-to-end with a new protocol. Doing this with a datagram protocol could be OK but is unless you do congestion control.
[16:31:14] <Gorry Fairhurst> David Black: This is not safe. A path that signals loss needs to react by doing Congestion Control.
[16:32:10] <Gorry Fairhurst> End of session
[16:32:16] Greg White leaves the room
[16:32:16] David Millman leaves the room
[16:32:17] Meetecho leaves the room
[16:34:29] Gorry Fairhurst leaves the room
[16:51:51] Gorry Fairhurst joins the room
[16:51:59] Gorry Fairhurst leaves the room
[17:41:48] scribeJohn leaves the room