[06:26:43] jlcJohn joins the room
[06:41:03] SwedeMike joins the room
[06:55:14] coopdanger joins the room
[06:55:24] wesley.m.eddy joins the room
[06:57:33] coopdanger leaves the room
[06:58:36] Matt Zekauskas joins the room
[07:01:42] Simon Leinen joins the room
[07:02:52] coopdanger joins the room
[07:03:32] Andrew McGregor joins the room
[07:03:51] swb joins the room
[07:06:53] <coopdanger> where are you guys in the agenda? (I'm in another WG so don't have audio)
[07:07:04] gigix73 joins the room
[07:07:08] <Andrew McGregor> Still doing the activity summary
[07:07:17] <coopdanger> thanks
[07:07:35] <Matt Zekauskas> any one in particular you're looking for?
[07:07:42] Jean-Francois M. joins the room
[07:07:53] <coopdanger> would like to come down for Jim Gettys
[07:08:03] <Matt Zekauskas> we'll try to give you a heads-up
[07:08:14] <coopdanger> thank you
[07:08:22] <swb> just got to 'office hours'
[07:09:12] <coopdanger> heading down now, saw what I needed to see in my other wg, thanks
[07:09:16] <Matt Zekauskas> k
[07:09:33] coopdanger leaves the room
[07:09:39] <Matt Zekauskas> (I'm trying to take notes, so wont' say much here)
[07:10:38] <Matt Zekauskas> Attempting to move to ALTO/DECADE presentation, I believe
[07:12:22] sdstrowes joins the room
[07:13:19] coopdanger joins the room
[07:14:30] rvdp joins the room
[07:16:05] sal joins the room
[07:20:14] <Matt Zekauskas> Starting bufferbloat
[07:21:30] <jlcJohn> Slide numbers would help...
[07:21:56] <jlcJohn> Oops... there aren't any on-slide, are there. :^(
[07:24:06] Shane Amante joins the room
[07:24:38] sal leaves the room: Replaced by new connection
[07:28:28] <Simon Leinen> No slide numbers, no :-(
[07:28:38] <Simon Leinen> "Triggers - saturation of the path" (slide title)
[07:29:18] <Simon Leinen> "What happens when a network is slow due to bufferbloat? - ..."
[07:30:24] <Simon Leinen> "Netalyzr: Illuminating Edge Network Neutrality..."
[07:31:52] <Simon Leinen> "Paranoia sets in"
[07:32:20] <Simon Leinen> "The Plot Thickens....... Home router experiments"
[07:32:41] <Simon Leinen> "During my tests, I thought I caught my home router..."
[07:33:12] <Simon Leinen> "Home router bufferbloat led me to host bufferbloat"
[07:34:09] Thomas Roessler joins the room
[07:35:03] <Simon Leinen> "Host bufferbloat, and your home router"
[07:36:15] <Simon Leinen> "But you may say, why don't I see bufferbloat on ethernet?"
[07:37:27] <Simon Leinen> "Where does host bufferbloat hurt?"
[07:37:54] <Simon Leinen> "Aggregate network behavoir - Back to the Future"
[07:39:17] <Simon Leinen> "Conversations with Van Jacobson"
[07:40:58] <Simon Leinen> "Aggregate bufferbloat locations"
[07:41:53] <Simon Leinen> "Fat Subnets: e.g. 802.11 or many other wireless technologies"
[07:42:48] <Simon Leinen> "What do you get? An ugly sight"
[07:43:39] <Simon Leinen> "Buffers are only detectable when they are *next* to ..."
[07:44:33] coopdanger leaves the room
[07:45:34] <Simon Leinen> "Other Locations of Bufferbloat"
[07:47:16] <Simon Leinen> "Where do bloated buffers hide?"
[07:47:54] <Simon Leinen> "Web Browser/servers - TCP initial congestion window changes"
[07:50:24] <Simon Leinen> "Network Neutrality Implications of Bufferbloat"
[07:51:41] <Simon Leinen> "What should the IETF do?
[07:52:20] <Simon Leinen> Time for questions!
[07:52:33] <Simon Leinen> Eliot Lear at the mike
[07:52:39] Jean-Francois M. leaves the room: Computer went to sleep
[07:53:08] <Simon Leinen> Q: What gives carriers' own telephony an edge?
[07:53:25] <Simon Leinen> A: e.g. in FiOS, telephony is provisioned on a completely separate pipe
[07:53:44] <Simon Leinen> Q (?, Google): Typically our interfaces have 150-200ms buffer capacity
[07:53:58] <Simon Leinen> (points to Curtis Villamizar's rules of thumb)
[07:54:32] <Simon Leinen> Nick (McKeown?)'s work on buffer sizing is actually overly pessimistic, buffers could be even smaller these days.
[07:55:02] <Simon Leinen> A: Buffer sizes often come out of thin air... it's very difficult to know enough about the environment (traffic, RTTs etc.) to dimension queues correctly
[07:55:16] <Simon Leinen> Google person: I have vendors now coming to me offering me buffer sizes up to one second
[07:55:55] <Simon Leinen> Q (Stuart Cheshire, Apple): I wrote something along these lines
[07:56:03] <Simon Leinen> (it's the latency, stupid-- SL)
[07:56:22] <Simon Leinen> One of your calls to action is to replace HTTP/1.1 - gues s the later preso on SPDY will address some of this.
[07:56:31] <Simon Leinen> About Active Queue Management (AQM)...
[07:56:56] <Simon Leinen> A: AQM is the ultimate solution to this. Unfortunately, RED is fundamentally flawed... Van believes it can't work in typical home networks
[07:57:06] <Simon Leinen> We need help from the real experts here.
[07:57:32] <Simon Leinen> We have to do queue management - even inside hosts
[07:58:01] <Simon Leinen> Stuart: I've been making this same argument within Apple... we also make base stations, but it's a small market share - have to get others on board
[07:58:21] <Simon Leinen> Mark Nottingham: Tomorrow at ?? meeting we'll have a presentation about pipelining(?) and also about SPDY
[07:58:49] <Simon Leinen> A: Pipelining was intended as a solution, unfortunately it seems very hard to implement.
[07:59:15] <Simon Leinen> Mark: Developers will always want to send many packets very quickly
[07:59:50] <Simon Leinen> Congestion control has always been some sort of "gentlemen's agreement"
[08:00:08] <Simon Leinen> Q: Joel Jaggli: I do need my large buffers on my 10G interface
[08:00:49] sal joins the room
[08:00:50] <Simon Leinen> From the datacenter: We build apps that manage to move 6.5 Gb/s over a path with 2s RTT - works fine for them.
[08:01:06] <Simon Leinen> Still Joel: We have really big flows that are really hard to shape...
[08:01:16] <Simon Leinen> IPSEC flows of multiple Gb/s
[08:01:56] <Simon Leinen> A: Whether a buffer gets you is whether it's adjacent to the congested link
[08:02:06] <Simon Leinen> Joel: It's also about where you can do something
[08:03:07] <Simon Leinen> Q: Matthew Kaufman: Very happy this is being brought up... about 16 years overdue. Was happy when Stuart brought this up 12 years ago - thought it would change something but it didn't. 10:04:23 AM
[08:03:25] <Simon Leinen> Vendors don't want to hear that they should build smaller buffers.
[08:03:37] <Simon Leinen> Hope that this time around we'll actually get the vendors to do something.
[08:04:04] <Simon Leinen> Jim: While AQM is probably the solution, let's not wait for this - there's things we can do to mitigate many cases where this hurts.
[08:04:34] <Simon Leinen> Examples: Buffers that are sized for 100Mb/s, but are used at 10-20 Mb/s...
[08:04:54] <Simon Leinen> Q: Colin Perkins: AQM and ECN are clearly Good Things to have
[08:05:18] <Simon Leinen> But just reducing buffers in the network might be more worthwhile
[08:05:39] <Simon Leinen> A: I think AQM is the real solution, but let's focus on what we can do.
[08:05:59] <Simon Leinen> OSes are now able to much better saturate typical links
[08:06:11] <Simon Leinen> Q: Joe Touch (ISI), we'll talk offline about the other 200 reasons for latency
[08:06:27] <Simon Leinen> Look at an ??? paper from 1997 about the hazards of using small buffers
[08:06:42] <Simon Leinen> Example of many people trying to use TELNET over a transatlantic link
[08:07:00] <Simon Leinen> Don't just reduce the buffers - send SHORTER packets!
[08:07:11] <Simon Leinen> ATM *almost* got this right, except they segmented at the lower layer
[08:07:30] <Simon Leinen> Pipelining: BAD IDEA.  We only move the problem one layer up...
[08:08:05] <Simon Leinen> Q: Murari, Microsoft: Reasons for AQM to not be deployed? (?)
[08:08:13] <Simon Leinen> A: Hard to configure correctly...
[08:08:30] Roy joins the room
[08:09:04] <Simon Leinen> Murari: As we learned with DCTCP, the problem is not just AQM - how can we achieve line rate with very small buffers?
[08:09:21] <Simon Leinen> A: Current buffers are so big that conegstion control breaks - TCP not even stable anymore.
[08:09:40] <Simon Leinen> Murari: Modern TCPs have delay-based congestion response - maybe much can be done just from the endpoints.
[08:10:06] <Simon Leinen> Q: Mark Handley: Glad you correctly identified cause of the problem.
[08:10:11] sal leaves the room
[08:10:16] <Simon Leinen> Saw from the nice, graphs that Jim was running CUBIC
[08:10:43] <Simon Leinen> We were successful! We wanted to be able to fill high-bandwidth large-capacity paths with TCP...
[08:10:57] <Simon Leinen> (I meant high-delay)
[08:11:27] <Simon Leinen> Mark: Our algorithms weren't designed for these kinds of larg buffers
[08:11:37] <Simon Leinen> Jim: We need timely congestion response
[08:11:56] <Simon Leinen> Q: (?) About large buffers in end hosts. We also have more multiplexing going on in hosts...
[08:11:57] <Matt Zekauskas> anyone know who is at mic?
[08:12:40] <swb> Jana Iyengar
[08:12:59] Jean-Francois M. joins the room
[08:13:04] <Matt Zekauskas> thanks!
[08:13:18] <swb> (Janardhan - Franklin & Marshall College - purveyor of "minions" as tools for transport protocol experimentation)
[08:13:35] <Simon Leinen> Next up: SPDY, TCP, and the Single Connection Throttle
[08:13:49] <Simon Leinen> http://www.ietf.org/proceedings/80/slides/tsvarea-0.pdf
[08:14:09] <Simon Leinen> Again, no slide numbers :-(
[08:15:04] <Simon Leinen> SPDY used on all Google servers and on newer Chrome, but just within SSL
[08:16:29] <jlcJohn> I'm clueless which slide he's on
[08:16:36] <Simon Leinen> "A New Protocol?  What for?"
[08:16:43] <Simon Leinen> "State of the Web Page"
[08:17:03] <Simon Leinen> (everything before was introduction, on the title slide)
[08:17:12] <jlcJohn> Thx
[08:18:22] <Simon Leinen> "Quick SPDY Background"
[08:21:15] <Simon Leinen> "HTTP Connection Use Today"
[08:22:02] <Simon Leinen> "Reducing Upload Bytes"
[08:22:35] Thomas Roessler leaves the room
[08:22:38] <Simon Leinen> Q: Stuart Cheshire: This is 6 connections per target hosts, and you get 30 connections because they are to different hosts, correct?
[08:22:40] <Simon Leinen> A: Correct.
[08:23:17] <Simon Leinen> "Reducing Download Bytes"
[08:24:01] <Simon Leinen> "Reducing Total Packets"
[08:24:16] <Simon Leinen> "Increasing Parallelism"
[08:24:44] <Simon Leinen> "The Single Connection Throttle"
[08:25:13] <Simon Leinen> "Throttle #1: CWND"
[08:25:27] <Simon Leinen> (Initial CWND, IW)
[08:26:00] gigix73 leaves the room
[08:27:11] <Simon Leinen> "Throttle #1 CWND vs # connections"
[08:28:57] <Simon Leinen> "Throttle #2: Receive Windows"
[08:29:35] <Simon Leinen> "Throttle #2: Init rwnd"
[08:29:41] <Simon Leinen> "Throttle #3: Intermediaries"
[08:30:48] <Simon Leinen> "Throttle #4: Congestion Control"
[08:32:05] <Simon Leinen> "Too Obsessed With 1 Connection?"
[08:33:06] <Simon Leinen> "How Much Does A Handshake Cost?"
[08:34:47] <Simon Leinen> "What's Next?"
[08:35:53] <Simon Leinen> Time for questions.
[08:36:00] <Simon Leinen> Q: Joe Touch, ISI
[08:36:35] <Simon Leinen> You're upset when someone in the middle of the network changes your Window Scaling, but you can just change the IW against the spec - that's OK, right?
[08:37:01] <Simon Leinen> A: Specs can be changed... it's coming - we can debate it but it's coming.
[08:37:19] <Simon Leinen> Q: Jim Gettys: My issue is what you're doing to other applications that may be much more delay-sensitive
[08:38:03] <Simon Leinen> Q: (?) Some of your issues (e.g. head-of-line blocking) have already been solved in SCTP...
[08:38:19] <Simon Leinen> A: Yes, could fix some problems, but not all, and deployability is a real issue.
[08:38:29] <Simon Leinen> Also, application layer is separate from transport layer
[08:38:47] <Simon Leinen> Q: Chicken-and-egg problem... maybe if Google started using SCTP, things might change.
[08:39:08] <sdstrowes> +1 on that
[08:39:10] <Simon Leinen> Q: Stuart Cheshire, Apple: Good work, thanks. To slide #3 (State of the Web Page)
[08:39:33] <Simon Leinen> Like to hear more about proxies and pipelining (if we have a couple of minutes)
[08:39:42] <Simon Leinen> A: Pipelining was written up about 10 years ago
[08:39:48] <Simon Leinen> never successfully deployed in a large browser
[08:39:55] <Simon Leinen> primarily because of proxies
[08:39:59] <Simon Leinen> that would cause random timeouts
[08:40:10] <Simon Leinen> Browser vendors couldn't take that risk, because the benefits didn't justify it.
[08:40:31] <Simon Leinen> Also, to use Pipelining, you have to know that the server speaks HTTP/1.1
[08:40:37] <Simon Leinen> That's another round-trip for finding out
[08:41:06] <Simon Leinen> Browser could cache this per server, but there's no guarantee that the next access to the "same" server still speaks 1.1
[08:41:36] <Simon Leinen> Browsers had to do background checks whether pipelining could really be used
[08:41:45] <Simon Leinen> Stuart: I really hope we can make Pipelining work
[08:41:51] <Simon Leinen> Example from iTunes store
[08:42:07] <Simon Leinen> iTunes uses the same WebKit that browsers do
[08:42:37] <Simon Leinen> Pipelining really improves performance dramatically for this page
[08:42:53] <Simon Leinen> A: Research on Pipelining should continue, it's great.
[08:44:02] <Simon Leinen> Q: Mark Nottingham: People gave up too early when trying Pipelining
[08:44:13] <Simon Leinen> Becoming more performance-savvy recently
[08:44:28] <Simon Leinen> It's worth spending next year trying to improve HTTP so that it's good enough
[08:44:32] <Simon Leinen> including Pipelining
[08:45:06] <Simon Leinen> SPDY is nice, immediately ran and implemented it - was fun
[08:45:26] <Jean-Francois M.> thanks Simon for being a great scribe and taking those notes, much appreciated on this end.
[08:45:59] <Simon Leinen> Q: Matthew Kaufman: Exchange experience between this and Websockets
[08:46:10] <Simon Leinen> A: Unclear how this can be integrated
[08:46:18] bcheck joins the room
[08:46:31] <Simon Leinen> docs.google.com is using SPDY today for "Websockets-like" things
[08:46:55] <Simon Leinen> Explains "long poll" technique for server-to-client push
[08:47:03] <Simon Leinen> "Hanging GET"
[08:47:29] <Simon Leinen> Hanging GETs eat away at the quota of parallel connections (per hosts)
[08:47:32] <Simon Leinen> SPDY helps a lot here
[08:47:43] <Simon Leinen> Matthew: There's more overlap than may be apparent.
[08:47:55] <Simon Leinen> Q: Salvador ?, HYBI chari
[08:48:10] <Simon Leinen> Websockets doesn't currently have support for multiplexing
[08:48:20] <Simon Leinen> Based on feedback from SPDY folks, try to improve this.
[08:48:53] <Simon Leinen> Q: Eliot Lear: Two applications protocols with pipelining experience: IMAP and (with considerably less success) BEEP
[08:49:08] <Simon Leinen> Suggests looking at SPDY over BEEP
[08:49:37] <Simon Leinen> Other point: The advantage that proxies don't mess with SPDY might wither over time
[08:50:50] <Simon Leinen> A: (long interesting digression about encryption)
[08:51:08] <Simon Leinen> Eliot: Even when SSL is used, that doesn't mean people can't mess with it...
[08:51:14] <Simon Leinen> A: Yes but that's evil
[08:52:04] bcheck leaves the room
[08:52:06] <Simon Leinen> Mark about some optimizations concerning SSL: False Start, ...?
[08:52:36] bcheck joins the room
[08:52:49] <Simon Leinen> Mark: Separate SSL from the protocol - I found it very valuable for use between servers (no SSL needed here)
[08:53:11] <Simon Leinen> Next up: "A Storage Menagerie: NAS, SAN & the IETF"
[08:53:12] <Matt Zekauskas> who asked about the separtion?
[08:53:15] <Simon Leinen> http://www.ietf.org/proceedings/80/slides/tsvarea-2.pdf
[08:53:17] Matt Zekauskas can't type
[08:53:21] <Simon Leinen> Matt: Mark Nottingham, I think
[08:53:58] sdstrowes leaves the room
[08:54:03] <Simon Leinen> Speaker: David Black, storm WG co-chair
[08:54:27] <Simon Leinen> (Yay, numbered slides!)
[08:54:27] <Simon Leinen> #2
[08:55:29] <Simon Leinen> #3
[08:55:56] <jlcJohn> You have good eyesight! I missed the numbers...
[08:56:40] <Simon Leinen> David passed the mike to Brian Pawlowski, nfsv4 WG co-chair
[08:56:51] ꈲ joins the room
[08:56:59] bcheck leaves the room
[08:58:43] bcheck joins the room
[08:59:15] <Simon Leinen> #4
[08:59:22] <Simon Leinen> #5
[08:59:24] coopdanger joins the room
[09:00:03] sdstrowes joins the room
[09:00:04] bcheck leaves the room
[09:00:40] <Simon Leinen> #6
[09:00:46] Jean-Francois M. leaves the room: Computer went to sleep
[09:00:50] bcheck joins the room
[09:01:09] <Simon Leinen> back to #5
[09:01:17] <Simon Leinen> #6
[09:02:15] <Simon Leinen> #7
[09:03:09] <Simon Leinen> #8
[09:04:19] <Simon Leinen> #9
[09:05:23] <Simon Leinen> (Handed mike back to David)
[09:05:28] <Simon Leinen> #10
[09:06:24] <Simon Leinen> #12
[09:06:44] <Simon Leinen> #13
[09:07:12] <Simon Leinen> #14
[09:07:44] <Simon Leinen> (btw. the slide #s aren't visible on the projection screen, but they are on the PDF)
[09:08:04] <Simon Leinen>  #15
[09:08:42] <Simon Leinen> #16
[09:08:51] <Simon Leinen> #17
[09:10:06] <Simon Leinen> #18
[09:10:57] <Simon Leinen> #19
[09:12:10] <Simon Leinen> #20
[09:12:47] <Simon Leinen> #21
[09:13:59] <Simon Leinen> #22
[09:14:58] <Simon Leinen> #23
[09:15:15] <Simon Leinen>  #24
[09:15:39] tleva joins the room
[09:16:21] <Simon Leinen> #25
[09:18:53] <Simon Leinen> #26
[09:19:28] <Simon Leinen> #27
[09:19:31] <Simon Leinen> #28
[09:19:55] <Simon Leinen> #29
[09:20:24] <Simon Leinen> #30
[09:20:52] <Simon Leinen> #31
[09:21:06] <Simon Leinen> #32
[09:21:27] <Simon Leinen> #33
[09:22:29] <Simon Leinen> Time for questions
[09:22:37] <Simon Leinen> Q: Joe Touch, ISI: Two short questions
[09:22:45] <Simon Leinen> Why (at least) two different to paths?
[09:23:05] <Simon Leinen> A: Reason for doing it: Distance extension (storage replication)
[09:23:34] <Simon Leinen> Want to transparently bridge fiber channel devices by rewrapping in TCP/IP in the middle
[09:23:52] <Simon Leinen> Joe: I'm missing the point - could just use iSCSI over TCP/IP
[09:24:10] <Simon Leinen> Maybe talk about this offline
[09:24:22] <Simon Leinen> A: This has to do with existing implementation infrastructure
[09:24:32] <Simon Leinen> Q: So it's historical?
[09:24:33] <Simon Leinen> A: Yes
[09:25:12] <Simon Leinen> Joe's other Q: What did you mean by "everything runs over Ethernet"
[09:25:23] Andrew McGregor leaves the room
[09:25:30] <Simon Leinen> A: Observing what seems to be happening in the datacenter world
[09:25:42] <Simon Leinen> rack-scale consolidation @10GE
[09:25:57] tleva leaves the room
[09:26:04] coopdanger leaves the room
[09:26:10] rvdp leaves the room
[09:26:17] Andrew McGregor joins the room
[09:26:18] sdstrowes leaves the room
[09:26:35] <Simon Leinen> Q: ?: Have people ever used erasure coding for replication?
[09:27:00] bcheck leaves the room
[09:27:11] <Simon Leinen> A: There are people working on that. But the main use case (disaster recovery) wants all data ready for use quickly in another place. Doesn't really fit well with erasure coding.
[09:27:39] <Simon Leinen> Session closed.
[09:27:40] Andrew McGregor leaves the room
[09:27:41] Shane Amante leaves the room
[09:28:04] wesley.m.eddy leaves the room
[09:28:04] ꈲ leaves the room
[09:29:25] Simon Leinen leaves the room
[09:35:39] swb leaves the room: Disconnected.
[09:41:04] Matt Zekauskas leaves the room
[09:41:32] Matt Zekauskas joins the room
[09:43:25] ꈲ joins the room
[09:44:50] swb joins the room
[09:47:50] swb leaves the room: Disconnected.
[09:48:01] swb joins the room
[09:51:39] Roy leaves the room
[10:02:09] coopdanger joins the room
[10:14:13] Roy joins the room
[10:22:25] ꈲ leaves the room
[10:23:51] wesley.m.eddy joins the room
[10:26:35] wesley.m.eddy leaves the room
[10:27:35] coopdanger leaves the room
[10:27:51] simon.leinen joins the room
[10:29:49] coopdanger joins the room
[10:44:01] swb leaves the room: Replaced by new connection.
[10:44:02] swb joins the room
[10:47:24] Andrew McGregor joins the room
[10:48:16] simon.leinen leaves the room
[10:48:47] wesley.m.eddy joins the room
[10:49:28] swb leaves the room: Disconnected.
[10:50:35] wesley.m.eddy leaves the room
[10:51:09] wesley.m.eddy joins the room
[10:52:05] coopdanger leaves the room
[10:53:35] Matt Zekauskas leaves the room
[10:56:32] Andrew McGregor leaves the room
[10:57:36] coopdanger joins the room
[10:59:05] Andrew McGregor joins the room
[11:00:00] swb joins the room
[11:04:52] Roy leaves the room
[11:07:02] coopdanger leaves the room
[11:07:05] Matt Zekauskas joins the room
[11:09:30] swb leaves the room: Disconnected.
[11:09:39] swb joins the room
[11:10:55] Matt Zekauskas leaves the room
[11:11:00] wesley.m.eddy leaves the room
[11:20:41] swb leaves the room: Replaced by new connection.
[11:20:43] swb joins the room
[11:23:08] sal joins the room
[11:41:24] sal leaves the room: Replaced by new connection
[12:06:16] Andrew McGregor leaves the room
[12:07:54] Andrew McGregor joins the room
[12:23:43] swb leaves the room: Disconnected.
[12:23:53] swb joins the room
[12:41:08] swb leaves the room
[12:41:29] Shane Amante joins the room
[12:41:48] Shane Amante leaves the room
[12:59:21] Andrew McGregor leaves the room
[13:57:55] Andrew McGregor joins the room
[14:16:27] Andrew McGregor leaves the room
[17:38:13] jlcJohn leaves the room