IETF
lsr
lsr@jabber.ietf.org
Tuesday, November 6, 2018< ^ >
Room Configuration
Room Occupants

GMT+0
[01:47:35] M8NEGe4t joins the room
[01:48:17] Meetecho joins the room
[01:55:16] Henk Smit joins the room
[01:55:23] Xufeng Liu joins the room
[01:59:06] Mikael Abrahamsson joins the room
[01:59:26] <Mikael Abrahamsson> I'll be jabber scribe, if you want something relayed to mic, please say "to mic:"
[02:00:34] Jeffrey Haas joins the room
[02:01:05] <Mikael Abrahamsson> meetecho, can you please put camera on pink box?
[02:01:15] <Mikael Abrahamsson> aim it, i mean
[02:02:44] <Mikael Abrahamsson> Les Ginsberg at mic
[02:05:50] <Meetecho> Done!
[02:06:04] <Meetecho> Thanks for the heads up, sorry for missing the context change
[02:08:37] Aijun Wang joins the room
[02:10:14] Hirochika Asai joins the room
[02:16:26] <Mikael Abrahamsson> Naiming Shen at mic
[02:17:24] Susan Hares joins the room
[02:17:39] <Mikael Abrahamsson> Acee Lindem at mic
[02:19:15] <Mikael Abrahamsson> Antoni Przygienda at mic
[02:21:40] <Mikael Abrahamsson> Antoni Przygienda at mic
[02:23:27] <Mikael Abrahamsson> Naiming Shen at mic
[02:24:25] Suneesh Babu joins the room
[02:30:10] <Henk Smit> the window with slides is all black. I suppose a laptop got disconnected from the projector ? Or is it just me ?
[02:30:37] <Jeffrey Haas> The chairs are struggling to find the correct presentation. :-)
[02:31:11] <Henk Smit> yes, I got that. :) Just wondering. But it's ok now. 2nd screen is showing slides now.
[02:39:01] <Mikael Abrahamsson> Tony Li at mic
[02:40:35] <Henk Smit> who is at the mic now ?
[02:40:55] <Henk Smit> alvaro maybe ?
[02:41:17] <Mikael Abrahamsson> I hav eno idea, he won't show me his badge.
[02:41:23] <Mikael Abrahamsson> someone from Huawei
[02:41:57] <Henk Smit> alvaro retana is with Huawei now, I think. I haven't heard his voice in 20 years :-) but I think it's him
[02:42:20] <Mikael Abrahamsson> no, I do not think this is alvaro.
[02:46:17] <Mikael Abrahamsson> Christian Martin at mic
[02:47:18] <Mikael Abrahamsson> Huaimo Chen was the person from Huawei before
[02:47:30] <Henk Smit> ok
[02:57:31] <Mikael Abrahamsson> Les Ginsberg at mic
[02:57:40] <Henk Smit> the answer is: yes, the result is a minimal spanning tree
[02:58:10] <Susan Hares> How does the algorithm prevent missed LSPs in the switch over case?
[02:58:34] <Henk Smit> combining drafts is combining a distributed and a centralized algorithm. so that might not be easy
[02:58:41] <Mikael Abrahamsson> Tony Li at mic
[02:58:56] <Henk Smit> Susan: when you change flooding state of an adjacency, you need to exchange CSNPs
[02:59:50] <Mikael Abrahamsson> Chris Bowers at mic
[03:00:39] <Mikael Abrahamsson> Christian Martin at ic
[03:05:57] Krishnaswamy Ananthamurthy joins the room
[03:08:39] Martin Horneffer joins the room
[03:17:29] Krishnaswamy Ananthamurthy leaves the room
[03:17:56] <Susan Hares> Henk was that in Gunter's document rather than his presentation?
[03:23:06] <Henk Smit> Yeah, using CSNPs is in the draft (I think). It certainly is what we propose. It wasn't in the presentation because it had to be short, and I think everyone needs to use CSNPs when the flooding-topology changes.
[03:23:13] <Mikael Abrahamsson> Antoni Przygienda at mic
[03:23:47] <Susan Hares> Henk - thanks - I will read the draft in detail.
[03:25:32] <Mikael Abrahamsson> Alvaro Retana at mic
[03:30:10] Aijun Wang leaves the room
[03:33:44] Aijun Wang joins the room
[03:36:54] <Mikael Abrahamsson> Antoni Przygienda at mic
[03:37:14] <Mikael Abrahamsson> Jeff Tantsura at mic
[03:39:07] <Mikael Abrahamsson> Sradhaat mic
[03:39:12] <Mikael Abrahamsson> Sradha at mic
[03:40:23] <Jeffrey Haas> Shraddha Hegde
[03:40:42] <Mikael Abrahamsson> thanks, she wasn't wearing a badge
[03:46:33] <Jeffrey Haas> To queue a mic question for end of presentation: "How does this proposal deal with TCP backpressure or window congestion that can lead to incomplete packets being queued in the stream"
[03:47:14] <Henk Smit> Not sure I understand. Isn't that the same as bgp update messages in a BGP TCP session ?
[03:47:31] <Mikael Abrahamsson> jeffrey: I think the header will take care of that, you know when you got it all
[03:47:50] <Jeffrey Haas> you end up with zero space in window.  a partial packet is sent. it only decongests as part of getting acks through.
[03:47:53] <Mikael Abrahamsson> everything is wrapped so you know when you have the complete LSP
[03:48:03] <Henk Smit> The way to think about this: if TCP gets congested, than the amount of "native" ISIS packets would have killed you for sure
[03:48:18] <Susan Hares> Queue a mic  question - are you proposing TCP or STCP?
[03:48:33] <Jeffrey Haas> tcp congests for weird issues all the time. it is what makes bgp "difficult"
[03:48:43] <Henk Smit> actually, each PDU (LSP, CSNP or initial IIH) has a length-field in its header. the marker is just to realize that you got a bug and are out of sync
[03:49:16] <Mikael Abrahamsson> ok, so chairs wanted this taken to the list
[03:49:27] <Jeffrey Haas> the issue isn't the incomplete packets.  it's the fact that IGPs currently deal with being in sync by having steady pacing of packets.
[03:50:11] <Henk Smit> I see your point. I can not comment how often that would happen in a real network.
[03:50:40] <Jeffrey Haas> speaking as someone who lives in bgp-land every day, tcp congestion is a regular headache.
[03:50:55] <Jeffrey Haas> bgp survives because each router does hop by hop
[03:50:55] <Henk Smit> Susan, we propose TCP. The TCP connection is always 1 hop. So I don't see a benefit of using STCP or something else.
[03:51:01] <Mikael Abrahamsson> BGP has keepalives to handle this, and consider the other end down (and session dead) after not receiving any packets for a while
[03:51:01] <Jeffrey Haas> igps are synchronized distributed map
[03:51:36] <Mikael Abrahamsson> jeffrey: but your worry is that you might get steady stream of packets but they are now minutes old?
[03:51:40] <Jeffrey Haas> Mikael, issue is not liveness.  you'd want to keep hello for that
[03:51:55] <Mikael Abrahamsson> jeffrey: so you can be behind and out of sync without realising?
[03:51:58] <Jeffrey Haas> not minutes old, simply delayed beyond when an spf gets triggered across the domain
[03:51:59] <Henk Smit> our draft suggests that the state of the TCP connection (blocked, silent, or broken) does not impact the state of the adjacency. that makes it easier. as long as IIHs are exchanged, the adjacency stays up
[03:52:04] <Jeffrey Haas> bingo - behind
[03:52:40] Martin Horneffer leaves the room
[03:52:57] <Jeffrey Haas> igps have a lot of mechanisms to deal with triggering the spf across the net.  these mechanisms minimally need tweaking in the presence of this potential - even if temporary - tcp congestion.
[03:53:37] <Jeffrey Haas> I'd strongly suggest the authors study tcp restransmit timer behaviors and also their interaction with even modest packet drop
[03:53:43] <Henk Smit> with X amount of LSPs being transmitted, do you think TCP would drop/hang/stall more than "native" ISIS PDUs being dropped ?
[03:54:12] <Jeffrey Haas> I agree with the core premise of the presentation: tcp can be significantly more productive for throughput.
[03:54:26] <Jeffrey Haas> however, issues related to what happens when congestion occurs must be dealt with
[03:54:40] <Mikael Abrahamsson> henk: I think the issue is that TCP could stall and sit on the packets minutes, trickling them out. packet-on-wire doesn't behave like that.
[03:54:46] <Jeffrey Haas> fwiw, my comments are core to my gripes in lsvr
[03:54:48] <Henk Smit> ok, thanks for the suggestion. I will certainly look into that.
[03:54:51] <Mikael Abrahamsson> Tonly Li at mic
[03:55:10] <Mikael Abrahamsson> Tony
[03:55:24] <Susan Hares> to mic - +1 to Tony Li due to problems.
[03:57:31] <Henk Smit> btw, my idea to use TCP for ISIS flooding happened when I did BMP (at my previous job). In a simul-environment, my BMP could send 100k or 200k msgs/sec without a problem. it was the receiver that slowed it down. so that made me think that with TCP (or Quic or whatever), we can speed up ISIS flooding over 1 adjacency by a large factor (compared to the 33 LSPs/sec as in 10589).
[03:58:35] Xufeng Liu leaves the room
[04:01:14] Meetecho leaves the room
[04:07:52] alessandro joins the room
[04:10:05] Simon Pietro Romano joins the room
[04:12:00] Simon Pietro Romano_7390 joins the room
[04:14:23] alessandro leaves the room
[04:14:33] Simon Pietro Romano_7390 leaves the room
[04:17:00] <Mikael Abrahamsson> TCP brings benefits but also potential problems, and I can see Jeffs concerns as valid
[04:17:40] <Mikael Abrahamsson> so doing something with its own congestion and loss detection (over UDP or something) might alleviate that. So that's where the QUIC suggestion comes from I think
[04:19:04] <Mikael Abrahamsson> also, how do you set up the IP connectivity? Is this done over link local per-hop?
[04:19:22] <Jeffrey Haas> It's been too long since I read the QUIC specs.  It fixes a lot of pacing and framing issues that makes TCP annoying
[04:19:40] <Jeffrey Haas> but I don't know that it completely deals with backpressure issues, especially for large IGP PDUs that need protocol fragmentation.
[04:19:50] <Mikael Abrahamsson> jeffrey: well, it conflates two levels so you always have insight into what's going on at the packet level
[04:20:28] <Jeffrey Haas> depends on your stack.  much for the same reason some BGP implementations are on custom-rolled TCP and some just use a kernel service.
[04:22:04] <Mikael Abrahamsson> jeffrey: I thought the major reason for separate TCP implementation was the TCP-AO?
[04:22:40] <Jeffrey Haas> no one implements (that I know of) tcp-ao yet.  security considerations are another very large discussion topic.
[04:22:54] <Mikael Abrahamsson> jeffrey: Right, I meant "TCP MD5"
[04:23:04] <Jeffrey Haas> tcp-md5 is integrated in even linux these days
[04:23:11] Meetecho joins the room
[04:23:43] <Mikael Abrahamsson> right, but several of these decisions are 15 years old I believe.
[04:23:53] <Mikael Abrahamsson> or 2
[04:23:56] <Mikael Abrahamsson> 20
[04:24:19] <Jeffrey Haas> I believe you've just summarized "why aren't we using a stream protocol to do our igp".  old decisions that made sense at the time.
[04:24:30] <Jeffrey Haas> understand why they made sense before casually changing them. :-)
[04:24:43] <Mikael Abrahamsson> right.
[04:25:38] <Henk Smit> Another inspiration for this draft is a discussion I had in 1996 with Dave Katz. Yakov and him were working on LDP. LDP uses TCP. Dave said: "if we had to do ISIS all over again, we'd probably use TCP now". I never forgot that. :)
[04:26:22] <Mikael Abrahamsson> well, the LDP choice of TCP actually has caused problems because it took longer for LDP to come up compared to ISIS, thus "IGP-LDP-SYNC" requirement
[04:26:25] <Mikael Abrahamsson> Les Ginsberg at mic
[04:28:32] <Mikael Abrahamsson> Antoni Przygienda at mic
[04:28:33] <Jeffrey Haas> arguably, we have similar headaches these days with BGP + BFD.  The BFD interactions are sometimes interop issues.
[04:30:36] <Mikael Abrahamsson> Stewart Bryant at mic
[04:31:28] <Mikael Abrahamsson> Les Ginsberg at mic
[04:36:30] <Henk Smit> didn't PNNI did something like this ?
[04:41:04] <Mikael Abrahamsson> Cengiz Alaettinpglu at mic
[04:41:19] <Mikael Abrahamsson> Cengiz Alaettinnglu at mic
[04:41:43] <Susan Hares> PNNI - something like this
[04:41:44] <Mikael Abrahamsson> Cengiz Alaettinoglu at mic
[04:42:34] <Mikael Abrahamsson> sorry, the correct spelling is "Cengiz Alaettinoglu"
[04:42:48] <Mikael Abrahamsson> Jeff Tantisura at mic
[04:43:59] <Mikael Abrahamsson> Christian Martin at mic
[04:44:46] Alessandro Amirante joins the room
[04:45:22] <Mikael Abrahamsson> Antoni Przygienda at mic
[04:45:34] Alessandro Amirante leaves the room
[04:46:01] <Mikael Abrahamsson> Christian Martin at mic
[04:46:02] Alessandro Amirante joins the room
[04:47:09] Alessandro Amirante leaves the room
[04:53:21] <Mikael Abrahamsson> Les Ginsberg at mic
[04:54:56] <Mikael Abrahamsson> Mikael Abrahamsson at mic
[04:55:40] Hirochika Asai leaves the room
[04:56:44] <Mikael Abrahamsson> Antoni Przygienda at mic
[04:57:50] <Mikael Abrahamsson> Jeff Tantisura at mic
[04:58:32] <Mikael Abrahamsson> Antoni Przygienda at mic
[04:59:12] <Mikael Abrahamsson> Les Ginsberg at mic
[05:00:35] <Mikael Abrahamsson> Jeff Tantisura at mic
[05:09:47] <Mikael Abrahamsson> Zafar Ali at mic
[05:11:08] Stanzin Takpa joins the room
[05:12:30] <Mikael Abrahamsson> I need to head out, so jabber scribe service by me is hereby terminated
[05:19:38] <Henk Smit> gn all. it's 6:19 am here. sleep well.
[05:20:28] Henk Smit leaves the room
[05:20:43] Meetecho leaves the room
[05:21:03] Mikael Abrahamsson leaves the room
[05:21:04] Suneesh Babu leaves the room
[05:21:04] Stanzin Takpa leaves the room
[05:21:04] Aijun Wang leaves the room
[05:21:04] Jeffrey Haas leaves the room
[05:21:04] Susan Hares leaves the room
[05:31:18] Simon Pietro Romano leaves the room
[06:38:53] Simon Pietro Romano joins the room
[06:55:55] Martin Horneffer joins the room
[07:03:03] Martin Horneffer leaves the room
[08:50:46] Simon Pietro Romano leaves the room