Displaying #traffic-server/2015-12-11.log:

Fri Dec 11 00:11:21 2015  biilmann:Joined the channel
Fri Dec 11 00:39:05 2015  biilmann:Joined the channel
Fri Dec 11 01:47:12 2015  yamc:Joined the channel
Fri Dec 11 01:55:38 2015  reveller1:Joined the channel
Fri Dec 11 02:00:28 2015  biilmann:Joined the channel
Fri Dec 11 03:02:10 2015  briang:bcall, it should
Fri Dec 11 03:02:34 2015  briang:the problem is the init_cwnd is only setup for the HttpClientSession
Fri Dec 11 03:02:40 2015  briang:not H2 or SpdySessions
Fri Dec 11 03:03:07 2015  briang:it also should be renamed because it contains "http_server_init_cwnd" or something like that, it should just be "net_init_cwnd" or somesuch
Fri Dec 11 03:24:46 2015  biilmann:Joined the channel
Fri Dec 11 04:54:20 2015  es:Joined the channel
Fri Dec 11 05:35:19 2015  jjr:Joined the channel
Fri Dec 11 05:47:46 2015  psp:Joined the channel
Fri Dec 11 07:16:48 2015  gancho:Joined the channel
Fri Dec 11 07:50:23 2015  biilmann:Joined the channel
Fri Dec 11 08:15:31 2015  slx86:Joined the channel
Fri Dec 11 08:36:29 2015  slx86:Joined the channel
Fri Dec 11 08:40:21 2015  slx86:Joined the channel
Fri Dec 11 08:44:26 2015  psp:Joined the channel
Fri Dec 11 08:54:40 2015  slx86:Joined the channel
Fri Dec 11 09:01:42 2015  slx86:Joined the channel
Fri Dec 11 09:09:03 2015  Lethalman:Joined the channel
Fri Dec 11 09:33:44 2015  lrea:Joined the channel
Fri Dec 11 09:41:02 2015  mturk:Joined the channel
Fri Dec 11 09:41:03 2015  mturk:Joined the channel
Fri Dec 11 11:31:01 2015  mturk:Joined the channel
Fri Dec 11 11:31:02 2015  mturk:Joined the channel
Fri Dec 11 12:48:39 2015  JSeymour:Joined the channel
Fri Dec 11 13:12:01 2015  gancho:Joined the channel
Fri Dec 11 13:19:07 2015  bahumbug:Joined the channel
Fri Dec 11 13:19:14 2015  niq:Joined the channel
Fri Dec 11 14:25:48 2015  niq:Joined the channel
Fri Dec 11 14:37:08 2015  esproul:Joined the channel
Fri Dec 11 14:38:14 2015  yamc:Joined the channel
Fri Dec 11 15:19:10 2015  zwoop:briang ? On Linux, you can't change the init cwnd per socket. It's done per route.
Fri Dec 11 15:19:49 2015  zwoop:so, it should work for any session, as long as you do it right on e.g. the default route
Fri Dec 11 15:23:39 2015  briang:zwoop, that's not entirely true.
Fri Dec 11 15:23:45 2015  zwoop:yeah ?
Fri Dec 11 15:23:51 2015  briang:1 sec
Fri Dec 11 15:24:24 2015  briang:actually you may be right
Fri Dec 11 15:24:32 2015  briang:but you _can_ change the congestion control algorithm per socket
Fri Dec 11 15:24:34 2015  briang:I know that for sure
Fri Dec 11 15:24:45 2015  briang:because I just read it in the man page for tcp(7)
Fri Dec 11 15:24:57 2015  zwoop:yeah, you can do that
Fri Dec 11 15:25:12 2015  briang:which is arguably more valuable for ATS anyway
Fri Dec 11 15:25:13 2015  zwoop:that might be new'ish ?
Fri Dec 11 15:25:18 2015  briang:nope
Fri Dec 11 15:25:20 2015  zwoop:yeah, maybe
Fri Dec 11 15:25:22 2015  briang:since Linux 2.6.13
Fri Dec 11 15:25:25 2015  zwoop:ah
Fri Dec 11 15:25:33 2015  zwoop:default init cwnd is 10 anyways on modern linux's
Fri Dec 11 15:25:38 2015  briang:right.
Fri Dec 11 15:25:51 2015  briang:but the problem we're running into is the one congestion control for everything at our edge
Fri Dec 11 15:25:59 2015  briang:it makes sense to use something like htcp on our backbone from our edge to our data center
Fri Dec 11 15:26:03 2015  briang:but to our users it doesn't
Fri Dec 11 15:26:09 2015  briang:they need something that handles packet loss better
Fri Dec 11 15:26:26 2015  briang:so we need to be able to do asymmetric congestion control
Fri Dec 11 15:26:39 2015  zwoop:makes sense
Fri Dec 11 15:26:53 2015  briang:the problem I'm running into is how to make it more granular than just one type for in and one type for all out sockets
Fri Dec 11 15:26:58 2015  briang:I want it to be changable once you get headers
Fri Dec 11 15:27:05 2015  briang:since it only really matters once you start sending data
Fri Dec 11 15:27:11 2015  briang:but that's turning out to be a nightmere w/ spdy and http/2
Fri Dec 11 15:27:13 2015  briang:easy with HTTP
Fri Dec 11 15:27:51 2015  briang:why would they let you change the congestion control per socket and not the init cwnd, that seems stupid
Fri Dec 11 15:27:57 2015  briang:linux is a weird world.
Fri Dec 11 15:28:01 2015  zwoop:ah
Fri Dec 11 15:28:21 2015  zwoop:yeah, google had a patch for the per socket init cwnd, but Linus -1'd it
Fri Dec 11 15:28:39 2015  briang:I wonder why...
Fri Dec 11 15:29:27 2015  briang:I wonder why no one has asked for this feature in ATS before
Fri Dec 11 15:29:33 2015  briang:it seems so valuable for reverse proxies.
Fri Dec 11 15:29:57 2015  zwoop:makes you wonder, you sure there's no way to do it per route as well ?
Fri Dec 11 15:34:09 2015  briang:zwoop, there is a way per-route
Fri Dec 11 15:34:30 2015  briang:using ip route ... initcwnd 10
Fri Dec 11 15:34:39 2015  zwoop:no, I mean for congestion control algorithm
Fri Dec 11 15:34:45 2015  briang:it's super easy to hard code per route
Fri Dec 11 15:35:03 2015  zwoop:that's not good enough for ya ?
Fri Dec 11 15:36:09 2015  briang:no because we don't have those routes
Fri Dec 11 15:36:18 2015  briang:for example for one user on a cellular asn I want to give them westwood
Fri Dec 11 15:36:21 2015  briang:and other users cubic
Fri Dec 11 15:36:31 2015  briang:that'd be super hard to do via linux ip routes, right?
Fri Dec 11 15:36:42 2015  briang:I want to do a lookup on the fly using an in memory geo ip db
Fri Dec 11 15:36:53 2015  briang:and then choose what is best based on the characteristics of that user.
Fri Dec 11 15:37:01 2015  briang:or sorry, the characteristics of that ASN
Fri Dec 11 15:37:05 2015  zwoop:well, if you know it per AS, you could certainly use bird or something to get the BGP routes from the router, right? And update the linux route config accordingly ?
Fri Dec 11 15:37:28 2015  zwoop:how would you know it inside of ATS?
Fri Dec 11 15:37:30 2015  zwoop:geo IP ?
Fri Dec 11 15:37:32 2015  briang:I know the IP
Fri Dec 11 15:37:40 2015  briang:and we already attach geo ip headers when the request arrives
Fri Dec 11 15:37:55 2015  zwoop:and you have some sort of DB with IPs that need a different algorithm?
Fri Dec 11 15:37:57 2015  briang:so technically for our case I just need to look at a header and the change congestion control before I start sending data.
Fri Dec 11 15:38:04 2015  zwoop:that seems really unwieldy to manage :/
Fri Dec 11 15:38:15 2015  briang:well it's probably gonna be something like:
Fri Dec 11 15:38:23 2015  briang:if mobile asn -> westwood, else cubic
Fri Dec 11 15:38:35 2015  briang:in reality I can't imagine we'll get too crazy
Fri Dec 11 15:38:57 2015  zwoop:right, so if yoiu know what ASNs are mobile, you could probe your routers to get those routes, no ?
Fri Dec 11 15:38:59 2015  briang:but we'll use our existing geo ip database however we can to determine if a user might benefit from an alternative congestion control.
Fri Dec 11 15:39:08 2015  briang:hmm.
Fri Dec 11 15:39:14 2015  briang:you mean all the ip4 space of that ASN?
Fri Dec 11 15:39:23 2015  briang:that I suppose would be doable
Fri Dec 11 15:39:24 2015  zwoop:right
Fri Dec 11 15:39:41 2015  zwoop:but yeah, seems klunky too
Fri Dec 11 15:39:42 2015  briang:yah that would work.
Fri Dec 11 15:39:54 2015  briang:that's a very ops approach to the problem lol
Fri Dec 11 15:39:57 2015  briang:I wanna write code :/
Fri Dec 11 15:39:57 2015  zwoop: :)
Fri Dec 11 15:40:01 2015  zwoop:haha
Fri Dec 11 15:42:19 2015  briang:I just finished setting up my own personal PoP in HK
Fri Dec 11 15:42:29 2015  zwoop:sweet
Fri Dec 11 15:42:49 2015  zwoop:if only we could get some geo-SLB for docs.trafficserver.a.o we could host our docs there too ;)
Fri Dec 11 15:42:54 2015  briang:turns out microsoft has a datacenter in HK for azure and as msdn subscribers (thanks apache) I get 150/mo free credit
Fri Dec 11 15:43:18 2015  briang:why don't we ask ultra or dyn to sponsor us?
Fri Dec 11 15:43:50 2015  briang:but that'd be apache I guess, right?
Fri Dec 11 15:44:02 2015  briang:we can't have our own authoritative dns servers other than apache, rigth?
Fri Dec 11 15:44:51 2015  pquerna:you could ask for NS records on the subdomain
Fri Dec 11 15:45:13 2015  zwoop:yeah, I just haven't had the stamina to try to wrangle up someone to give it to us for free
Fri Dec 11 15:45:30 2015  zwoop:but PSUdaemon might help us here, and we can just purchase DNS services
Fri Dec 11 15:46:09 2015  briang:i seriously can't imagine that we'd get soo many queries that it'd really cost that much
Fri Dec 11 15:46:40 2015  zwoop:yeah, for sure
Fri Dec 11 15:46:46 2015  briang:we need to get an intern that can help redesign ts.apache.org
Fri Dec 11 15:46:51 2015  PSUdaemon:?
Fri Dec 11 15:46:54 2015  zwoop:there were some discussions that ASF infra would get DNS via Amazon I think
Fri Dec 11 15:47:12 2015  zwoop:PSUdaemon we could earmark some of your money to get services we need (such as a geo-SLB capable DNS service)
Fri Dec 11 15:47:27 2015  PSUdaemon:sure
Fri Dec 11 15:47:36 2015  zwoop:such that we can have docs.trafficserver.a.o in Europe, Asia and east / west US
Fri Dec 11 15:47:59 2015  briang:tbh, I think we'd be better off spending some money to have the site redesigned :p
Fri Dec 11 15:48:07 2015  zwoop:I'm sure the cost would be marginal, like < $100 / year
Fri Dec 11 15:48:13 2015  briang:as much as i'd love to see docs.ts faster
Fri Dec 11 15:48:14 2015  zwoop:briang we could do that too
Fri Dec 11 15:48:35 2015  briang:docs are awesome now let's rebrand ATS w/ a sexy new bootstrap template haha
Fri Dec 11 15:48:53 2015  zwoop:I sincerely dislike RTD, because a) it's impossible to setup your own RTD instance and b) it's impossible otherwise to get it to use the proxy for static content (the idiots use absolute URLs)
Fri Dec 11 15:50:05 2015  briang:but it does look nice ;/
Fri Dec 11 15:50:33 2015  briang:the other option we have is we could write some scripts that "mirror" it by making a full copy of the entire read the docs site every day
Fri Dec 11 15:50:38 2015  briang:and then just managing our own mirror
Fri Dec 11 15:50:48 2015  zwoop:briang we've spent a small fortune paying OmniOS to work on the Docs site
Fri Dec 11 15:50:52 2015  zwoop:(we == Apple)
Fri Dec 11 15:51:07 2015  zwoop:OmniTI ...
Fri Dec 11 15:51:44 2015  zwoop:The money we *maybe* get from PSUdaemon would let us develop like 2 pages on the main site, it's hardly worth it wasting it on that.
Fri Dec 11 15:52:29 2015  zwoop:But if we can spend $100 to make the site(s) faster, I think it's worth it
Fri Dec 11 15:52:37 2015  zwoop:we can use it for docs, ci and maybe even the main page
Fri Dec 11 15:52:58 2015  zwoop:The wiki would be a lot more difficult to put behind ATS as it is now.
Fri Dec 11 15:53:10 2015  briang:how much do you think it would cost to redesign the site?
Fri Dec 11 15:53:17 2015  briang:I have no idea how much these things cost
Fri Dec 11 15:53:27 2015  zwoop:depends
Fri Dec 11 15:53:33 2015  zwoop:i'd imaigne no less than $100 / hour ?
Fri Dec 11 15:53:39 2015  zwoop:(which seems cheap)
Fri Dec 11 15:53:48 2015  zwoop:imagine at least 20-40 hours of work ?
Fri Dec 11 15:54:20 2015  briang:2-3k doesn't seem bad
Fri Dec 11 15:54:45 2015  zwoop:that'd be a huge chunk of the money that we could otherwise use for e.g. lunch at summits
Fri Dec 11 15:55:29 2015  zwoop:I dunno if you can get someone (good) to do this work for $100 / hour though :). Also, if we do this, we should get rid of the insane ASF CMS nonsense.
Fri Dec 11 15:55:31 2015  briang:yah I like lunch ;/
Fri Dec 11 15:55:34 2015  briang:so that's a dilema
Fri Dec 11 15:55:55 2015  zwoop:well, all this is moot until PSUdaemon figure his shit out :)
Fri Dec 11 15:56:06 2015  briang:confused where PSUdaemon comes into it
Fri Dec 11 15:56:16 2015  zwoop:he's a baller
Fri Dec 11 15:56:20 2015  briang:does he has css skills?
Fri Dec 11 15:56:32 2015  briang:because that also solves the problem haha
Fri Dec 11 15:56:40 2015  zwoop:lol
Fri Dec 11 15:57:32 2015  zwoop:I'm sure he'll be all over it
Fri Dec 11 15:59:26 2015  PSUdaemon:i'm not sure we would be allowed to spend money on a site redesign
Fri Dec 11 15:59:35 2015  PSUdaemon:that might be considered "paying for code"
Fri Dec 11 16:00:19 2015  PSUdaemon:briang: i have no skillz
Fri Dec 11 16:00:23 2015  PSUdaemon:other than drinking
Fri Dec 11 16:00:30 2015  jumby:Joined the channel
Fri Dec 11 16:10:45 2015  dxu:Joined the channel
Fri Dec 11 16:29:18 2015  lrea:Left the channel
Fri Dec 11 16:53:30 2015  jpeach:Changed the topic to: Apache Traffic Server | https://docs.trafficserver.apache.org | Current release: 6.0.0 | Channel logs at http://wilderness.apache.org/channels/?c=traffic-server
Fri Dec 11 17:22:03 2015  biilmann:Joined the channel
Fri Dec 11 17:35:18 2015  jpeach:PSUdaemon: why did you add the length parameter to lookup() and not use it?
Fri Dec 11 17:36:10 2015  gancho:Joined the channel
Fri Dec 11 17:36:13 2015  PSUdaemon:likely it was used at some point and then later not
Fri Dec 11 17:36:30 2015  PSUdaemon:probably could be cleaned up
Fri Dec 11 17:36:47 2015  PSUdaemon:but i think johns code is where that should happen
Fri Dec 11 17:37:06 2015  PSUdaemon:this is to get some bugs that we have dealt with in our own branch upstream
Fri Dec 11 17:37:20 2015  PSUdaemon:this is the accumulation of a lot of small commits over time
Fri Dec 11 17:37:28 2015  jpeach:no John's code doesn't use it either
Fri Dec 11 17:37:37 2015  PSUdaemon:probably my bad for not being more on top of getting this stuff upstream faster
Fri Dec 11 17:38:35 2015  jpeach:if it is not needed can we remove it?
Fri Dec 11 17:38:45 2015  PSUdaemon:sure
Fri Dec 11 17:38:53 2015  PSUdaemon:let me check git history real quick
Fri Dec 11 17:38:58 2015  PSUdaemon:see where it entered
Fri Dec 11 17:39:55 2015  PSUdaemon:where exactly is this?
Fri Dec 11 17:40:07 2015  PSUdaemon:ConsistentHash.cc?
Fri Dec 11 17:40:13 2015  PSUdaemon:line 70?
Fri Dec 11 17:40:44 2015  jpeach:yeh
Fri Dec 11 17:41:03 2015  PSUdaemon:it's used on like 104
Fri Dec 11 17:41:20 2015  PSUdaemon:or are you saying i never call it with a len?
Fri Dec 11 17:41:43 2015  jpeach:you never call lookup() or lookup_available() with a len
Fri Dec 11 17:42:22 2015  PSUdaemon:k
Fri Dec 11 17:42:27 2015  PSUdaemon:i think i know why that is
Fri Dec 11 17:42:32 2015  PSUdaemon:and i can remove it
Fri Dec 11 17:42:42 2015  PSUdaemon:i think there was an evolution with the path hash
Fri Dec 11 17:43:58 2015  jpeach:thx ... it struck me as odd when I looked at rushford's PR becuase I could not see it being used
Fri Dec 11 17:47:18 2015  PSUdaemon:i think the first attempt was made where we igored qstring by passing in a length that did not include the qstring
Fri Dec 11 17:47:46 2015  jpeach:ah I see
Fri Dec 11 17:48:55 2015  gancho:Joined the channel
Fri Dec 11 17:49:48 2015  PSUdaemon:jpeach: zwoop: would you guys like me to rebase these altogether or keep seaprate commits?
Fri Dec 11 17:51:01 2015  jpeach:I would squash removing the url_len parameter
Fri Dec 11 17:55:39 2015  PSUdaemon:jpeach: yeah, that looks cleaner now
Fri Dec 11 17:55:41 2015  PSUdaemon:thx
Fri Dec 11 18:01:26 2015  biilmann:Joined the channel
Fri Dec 11 18:07:08 2015  gancho:Joined the channel
Fri Dec 11 18:11:28 2015  es:Joined the channel
Fri Dec 11 18:21:28 2015  biilmann:Joined the channel
Fri Dec 11 18:34:16 2015  gancho:Joined the channel
Fri Dec 11 18:50:54 2015  gancho:Joined the channel
Fri Dec 11 19:14:43 2015  jrushford:Joined the channel
Fri Dec 11 19:22:07 2015  _klk_:Joined the channel
Fri Dec 11 19:38:34 2015  meeramn:Joined the channel
Fri Dec 11 19:57:09 2015  gancho:Joined the channel
Fri Dec 11 20:05:23 2015  biilmann:Joined the channel
Fri Dec 11 20:10:25 2015  _klk_:Left the channel
Fri Dec 11 20:18:14 2015  meeramn:Joined the channel
Fri Dec 11 20:23:55 2015  zwoop:jpeach http://pastebin.com/iEByyrxQ . Assuming this is correct, how do we get clang-analyzer not to harp over sizeof(e) ?
Fri Dec 11 20:25:00 2015  jpeach:what problem does it have?
Fri Dec 11 20:25:16 2015  zwoop:it doesn't like doing sizeof(e);
Fri Dec 11 20:25:26 2015  zwoop:where e is a pointer
Fri Dec 11 20:25:57 2015  jpeach:that can't be the underlying problem though
Fri Dec 11 20:26:19 2015  jpeach:does it think you ought to be doing sizeof(*e)?
Fri Dec 11 20:26:30 2015  dxu:Is the apache git repo down for anyone else? Tried cloning from 2 machines on 2 networks and they both failed
Fri Dec 11 20:27:33 2015  zwoop:jpeach yes
Fri Dec 11 20:27:53 2015  zwoop:"The code calls sizeof() on a pointer type. This can produce an unexpected result"
Fri Dec 11 20:28:09 2015  zwoop:but, I suspect it's right here
Fri Dec 11 20:28:28 2015  zwoop:but, not 100% sure
Fri Dec 11 20:28:41 2015  zwoop:I think these methods were added to support the new RAM cache regressions
Fri Dec 11 20:28:49 2015  jpeach:I guess it depends on what RamCacheLRU::size means
Fri Dec 11 20:29:15 2015  jpeach:I think it is right
Fri Dec 11 20:29:37 2015  zwoop:yeah
Fri Dec 11 20:29:51 2015  zwoop:so, how do we tell clang analyzer to stop being a dick ?
Fri Dec 11 20:29:52 2015  jpeach:ram cache sise should be sizeof(entry) + sizeof(IOBufferData) + sizeof(the data block)
Fri Dec 11 20:30:04 2015  jpeach:I'm saying the analyzer is right
Fri Dec 11 20:30:07 2015  zwoop:oh
Fri Dec 11 20:30:23 2015  jpeach:gotta step out for a couple of hours
Fri Dec 11 20:30:47 2015  jpeach:I think it is meant to be sizeof(struct RamCacheLRUEntry) because sizeof(struct RamCacheLRUEntry *) would not make sense here
Fri Dec 11 20:31:24 2015  zwoop:ok
Fri Dec 11 20:36:14 2015  zwoop:I'll ask John to confirm :)
Fri Dec 11 20:55:59 2015  gancho:Joined the channel
Fri Dec 11 21:44:12 2015  meeramn:Joined the channel
Fri Dec 11 22:14:22 2015  dustywusty_:Joined the channel
Fri Dec 11 22:29:16 2015  zwoop:TS-4070
Fri Dec 11 22:29:26 2015  zwoop:bcall this is interesting, might need a backport for this too
Fri Dec 11 22:52:39 2015  meeramn:Joined the channel
Fri Dec 11 22:56:53 2015  gancho:Joined the channel
Fri Dec 11 22:59:23 2015  _klk_:Joined the channel
Fri Dec 11 23:06:55 2015  phibs:Joined the channel
Fri Dec 11 23:07:55 2015  phibs:Joined the channel
Fri Dec 11 23:09:34 2015  phibs:Joined the channel
Fri Dec 11 23:10:30 2015  phibs:Joined the channel
Fri Dec 11 23:16:40 2015  _klk_:Left the channel
Fri Dec 11 23:16:59 2015  phibs:Joined the channel
Fri Dec 11 23:40:41 2015  phibs:Joined the channel
Fri Dec 11 23:56:12 2015  jrushford:Joined the channel
Fri Dec 11 23:59:57 2015  niq:Joined the channel

Comments