16:59:57 #startmeeting Network team meeting, 14th of june 2021 16:59:57 Meeting started Mon Jun 14 16:59:57 2021 UTC. The chair is ahf. Information about MeetBot at http://wiki.debian.org/MeetBot. 16:59:57 Useful Commands: #action #agreed #help #info #idea #link #topic. 17:00:03 hello hello everybody 17:00:11 oh hi! 17:00:13 pad is at https://pad.riseup.net/p/tor-netteam-2021.1-keep 17:00:13 o/ 17:00:19 * nickm does ceremonial meeting dance 17:00:26 we are missing david today who is out 17:00:45 o/ 17:01:29 how are folks gitlab dashboards looking? 17:01:43 I'm happy with mine 17:02:07 o/ 17:02:08 i am not super happy with mine but that is my own fault 17:02:10 all good here updated 17:02:14 excellent 17:03:29 our two queries for releases looks the same as last week, with exception of the leap second ticket having gotten some attention since last meeting 17:04:25 * ahf is considering closing tor#40185 but that can be done post-meeting 17:05:07 ok! 17:05:14 (ugh out of battery. relocating 3 mins brb) 17:05:19 good luck 17:05:33 i see nothing new from other team who needs something from us that ISNT being handled already 17:05:42 the s30 tickets are there and david seems to be on the NHT tickets 17:06:23 we have a single announcement: today is the day where i enable 2fa enforcement for /tpo/core/** membership on gitlab, but i think everybody have enabled it now 17:06:32 if not, now is a great time 17:06:49 that is all our mandatory meeting agenda items. i think mikeperry have some s61 stuff for us to talk about 17:06:51 {also: reminder that i'm doing reduced hours this month; see tor-internal post} 17:06:58 ya! 17:07:21 yes, very exciting news 17:07:45 I went down a congestion control rabbithole last week and into the weekend and implemented 3 congestion control algs 17:08:01 I tested them with onion services, and all out-perform SENDME on the live network 17:08:27 by how much? 17:08:32 (I don't advocate or endorse working on weekends, but the live testing meant I could find and fix issues quickly and just got into an obsessive loop) 17:08:45 I will probably take a day off this week 17:08:52 sweet! 17:09:14 Oh: Reminder that Friday we're observing a US holiday. 17:09:22 and yes, please do - is it something people can start toying around with already to build up some understanding of these code changes? 17:09:25 oh shit, I didn't know 17:09:29 thanks 17:09:36 i assume in the end we may include multiple CC algorithms in tor and let the consensus pick the one we prefer? 17:10:27 anyway the alg perf compared to sendme is hard to say because I was just using random circuits. I saw anywhere from 3-5X faster.. but because I was tuhning against live which is all SENME, I may have made them too agrtessive (which means they will over-queue when deployed en-masse) 17:10:36 we will still need shadow to learn that 17:11:18 exciting 17:11:23 ahf: yes, the plan is to tune them in shadow and get an idea on best algs and params there, and then verify on live once enough relays and clients have upgraded 17:11:32 very nice 17:11:45 mikeperry: nicee wrt congestion control algos 17:12:10 nod 17:12:19 I need to do a lot of refactoring and implement stream flow control still.. I'd guess maybe a couple weeks before ready for prelim review of code structure 17:13:07 (so we can review code structure and refactor per review feedback while we churn through endless shadow sims) 17:13:26 if you need relays on the production network some hax patches applied before, we can also do that to get some feeling with it there too 17:13:27 btw I chatted with Rob last week about modeling CPU latency; he agrees about the basic approach we discussed but we both think it'll be a pretty rough estimate. hopefully we'll end up showing that the CPU usage doesn't make much difference and we can ignore it 17:13:35 it will also be useful to set up an onion service test on live, with fixed paths 17:14:20 to further check live behavior, since this is a faster testing loop than shadow sims 17:14:56 I hacked this in my branch, but it can be made much more rigourous for better datapoints 17:15:14 makes sense. i have no idea how to pin the entire path for an OS though 17:15:39 yes I am hoping dgoulet and/or asn can help with that 17:16:08 it might also help if we pin some relays we run, so we can look at their queues 17:17:08 yeah 17:17:17 jnewsome: how about client models from rob's previous experiments? one thing he may not have done that we will need is an uploading client 17:17:37 because that is where the flow control behavior will matter 17:17:57 mikeperry: that should be an easy tgen config 17:18:28 mikeperry: perhaps a foolish question: do your tests verify that the correct data, and the correct amount of data is actually transferred? I've messed up in the past by only counting the time up to an EOF 17:18:41 actually one nice thing about shadow 2.0 is if you already have a specific client you want to test, we may be able to run it directly in shadow 17:19:11 nickm: I was just testing with a curl + onionshare. but good call I will check sha's and such 17:19:21 mikeperry: great 17:20:33 in general I want better metrics from the live tests than I have... need to gather RTT stats, and get better throughput infos 17:20:52 but I still learned a lot and fixed a lot of things from the basic testing 17:21:14 i forgot to say some things in the proposal too that I realized after re-reading my old posts :/ 17:21:23 so there will be proposal updates 17:22:11 cool 17:22:12 anw that's super exciting. 17:22:15 ya 17:22:42 we should probably make some comms noise too when we start having this in the tor code 17:22:58 also remember re perf: this is a network effect thing.. the perf increases I saw will be small compared to what happens when all clients switch 17:23:06 if we tune it right in shadow 17:24:01 I think that's it for congestion control, unless there are other questions 17:24:10 cool 17:24:15 GeKo,juga: did you want to talk about the flooding experiment and/or sbws? 17:24:30 i can 17:24:41 we found a critical bug in sbws 17:24:54 so we asked one dir auth op to move back to torflow 17:25:03 oof 17:25:06 until we have investigated and a fix 17:25:18 oh ok :-/ 17:25:22 sbs#40091 is the ticket 17:25:28 sbws#40091 17:25:34 it seems juga has something to test 17:25:36 hi 17:25:45 and i'll review it tomorrow 17:26:02 hopefully we can get the third bwauth back to sbws like next week 17:26:16 hihi 17:26:23 o/ 17:26:31 juga: anything you feel we should add here? 17:27:09 for the flooding part 17:27:21 i collected all the questions we gathered so far 17:27:26 hmm, not really, just that maybe tjr and i are finding the issue with sbws stalling and/or out of memory 17:27:34 and created a new ticket to not mess up the old one 17:27:37 https://gitlab.torproject.org/tpo/metrics/analysis/-/issues/40001 17:28:04 juga: right, that's one of the other sbws mystery's we try to figure out 17:28:19 it's only happening intermittently and only on one bwauth... 17:28:22 so, exciting :) 17:28:35 * juga will update #40092 with the findings 17:28:50 nice 17:28:51 i think that's all for those two items 17:29:26 * GeKo hands mic back to mikeperry 17:30:13 I think I am good unless there are more questions. also happy to get volunteers for help with live testing and instrumented relays 17:30:27 sad dgoulet is out, I am guessing he can help there tho 17:30:34 I have refactoring and cleanup to do anyway 17:30:54 i will happily help with the relay stuff and testing on live network. david also have shell on those boxes though if he dives into it faster 17:31:46 shall we call it then and get back to our non-irc shells 17:31:56 anything we are missing? 17:32:00 jnewsome: lmk once you get that docker thing going on the new box. it is a bit early to test this branch in shadow still - it is hacky. but soon! :) 17:32:28 mikeperry: only a note to say that anarcat already has all the info we need to setup the shadow server. lavamind will do it. 17:32:32 mikeperry: cool, still waiting on @anarcat to get access 17:32:40 who I believe is waiting on cymru for credentials 17:32:52 jnewsome: not waiting anymore. all set and it will happen this week. 17:33:10 gaba: woohoo! 17:33:33 i'm going to double-check that right now 17:34:14 confirmed, i do have some access 17:34:24 the setup of those boxes is tricky and might take a while 17:36:39 ok 17:36:43 i am gonna disable the bot now 17:36:48 talk to you all in the other channel(s) 17:36:50 #endmeeting