mozilla :: #taskcluster

17 Mar 2017
13:38dustinfwiw I suspect that those permanent creds will outlast BBB
13:39gerard-majaxdustin, hello there
13:39gerard-majaxdustin, I was wondering about something regarding taskcluster
13:39gerard-majaxdustin, I may want to create a debian chroot during a build, this relies on sudo + multistrap (like debootstrap)
13:39gerard-majaxdustin, is this something that could work?
13:40dustinif it can work in docker
13:40gerard-majaxI have no idea if it can :)
13:40dustinfor reasons I don't really understand, fedora's "mock" doesn't work in docker
13:40gerard-majaxI mostly worry about the sudo usage though
13:40dustinsudo isn't a problem
13:40gerard-majaxmultistrap requires root at some point, I can't recall why exactly
13:42jmaherwcosta: ted: interesting thought on changing the version of libc++
13:43wcostajmaher: was a comment in the bug? I am afk
13:43jmaherwcosta: oh yeah; ted just added it not too long ago
13:45wcostaCan you post the link to the bug? I think I found the issue, I am waiting the try to conclude
13:45firebotBug 1338651 NEW, taskcluster cross-compiled OS X builds show Talos performance regressions vs. buildbot builds
13:46jmaherwcosta: oh, found the issue?!?
13:48tedwcosta: whoa!
13:48dustingarbas: is something you can do?
13:48firebotBug 1348284 NEW, Rename shipit bot hook roles
13:48wcostajmaher garndt: looking at the assembly code again, I noticed that the tc version wasn't using 64 bits mode instructions and registers, then I found out bb uses target x86_64-darwin11 and tc x86_64-darwin
13:49dustinwcosta: wow, nice work!
13:49tedthat got changed at some point because...
13:49* ted tries to remember exactly
13:49tedwe were using the full kernel name there (with version) to locate the toolchain binaries, IIRC?
13:50garndtthat would be awesome
13:50* garndt is excited
13:50wcostated: in treeherder you can find a commit "fix cross compiling" or something that has the change
13:50tedwcosta: oh, hm
13:51tedwcosta: good catch, anyway!
13:51wcostayesterday I tried with only this fix, but still got regression, so today I am testing with other changes I made to see the results
13:52catleeyou think that explains the regressions?
13:53wcostacatlee: I guess so, definitely this is an issue, I am just unsure if it is the only issue
13:53garndtyea it seems unclear if there is just a silver bullet for everything, but perhaps chipping away at this reduces how much the regression is
13:54jmaherI would bet on a silver bullet for at least 60% of the regressions, hopefully 90%
13:54garndtthat would be great
13:54catleeso exciting
13:54tedwell, the regression seems to be entirely due to IPC overhead
13:54tedi could fully believe that it's some simple low-level thing
13:55tedlike i said in the bug, the code's not *broken*, as evidenced by the fact that the debug tests are green
13:55jmaherwasn't e10s supposed to make firefox faster? ;)
13:55wcostaI made other changes, like downgrade clang version, add some missing linker flags and so
13:56tedthe clang versions should definitely be the same for perf comparison
13:56wcostajmaher: yes, but getting ipc right is tricky
13:57wcostaI am confident that my past try push will show no performance regressions, then it will be a matter of finding out which patches that fixed it
14:00wcostajmaher: for an example, mconley has a series called "the joy of code", one episode shows an e10s regression (btw, with this video I learned to interpret cleopatra profiles)
14:00wcostaI can post the link when I am on computer
14:01garndtI really l ike the joy of code even if 99% of it is over my head
14:01jmaherwcosta: awesome; mconley your videos are valuable
14:22mconleyhooray having an impact
14:24mconleygarndt: cards on the table, a good chunk of it is over my head too. :) Last episode (Ep 95), I was investigating a bug and just getting nowhere. Thrashing like crazy.
14:25garndtI'll have to check that one out
19:05wlachdustin: review ping re: bug 1347698, should be a fast one :)
19:05firebot NEW, Add support for retriggering reftests in automation with extra debugging options
19:14wlachdustin: thanks! :)
18 Mar 2017
No messages
Last message: 154 days and 21 hours ago