mozilla :: #perf

17 Jul 2017
11:46jmaher|afkhi igoldan
11:47igoldanjmaher|afk: hi, Joel
11:47igoldanjmaher|afk: how are you doing?
11:47jmaher|afkigoldan: great, how are you doing
11:48igoldanjmaher|afk: doing some retriggers & backfills on platform microbench
11:48igoldanjmaher|afk: then I'll go through the tracking bugs
11:48jmaherigoldan: great
11:48igoldanjmaher|afk: then continue with the Talos maintanance tasks
11:49jmaherigoldan: we should really finish up before doing a lot of platform micro benchmarks
11:50igoldanjmaher: ok
11:52jmaherigoldan: it would at least make things easier :)
11:53jmaherigoldan: there was one regression I emailed about this weekend- it seems that many tests got noisy
11:53igoldanjmaher: I don't remember reading an email about this
11:58jmaherigoldan: aha, I sent it to your email, just forwarded it to you
11:58igoldanjmaher: oh, I see now
12:01igoldanjmaher: I noticed that win7 pgo issue
12:02igoldanjmaher: I'm thinking that maybe the build configuration for PGO was changed
12:02jmaherigoldan: yeah, I really want to figure this out, it seems to make many tests invalid
12:02jmaherigoldan: that would be a likely explanation- or some other build flag that triggers when pgo is built
12:03igoldanjmaher: yes, this 2nd option is plausible
12:03igoldanjmaher: and I could bring a confirmation/or not for it, by looking into the pushes
12:04igoldanjmaher: as soon as I reach a conclusion, I'll let you know
12:05igoldanjmaher: before that, I do have some questions regarding this task:
12:05igoldanjmaher: how do you test Talos changes locally?
12:06igoldanjmaher: did you prepare a special virtualenv for Python2.7?
12:06igoldanjmaher: did you make use of the script?
12:07igoldanjmaher: I want to know that we use the same configuration when changing the Talos harness
12:07jmaherigoldan: ./mach talos-test -a <test>
12:07jmahertalos-test creates the virtualenv
12:07igoldanjmaher: and that virtualenv has Talos installed in edit mode, I guess
12:08jmaherigoldan: yes, but you need to pay attention to the paths if you are going to reuse it
12:10igoldanjmaher: thanks for telling this
13:16igoldanjmaher: I&#39;m having some difficulties when retriggering for Autophone alerts
13:16igoldanjmaher: I mean: from Treeherder, I see multiple values for a test
13:17igoldanjmaher: but on the graphs, I only see one data point, for the same changeset
13:20jmaherigoldan: please be very careful with autophone retriggers
13:21jmaherigoldan: we have very limited resources there
13:21jmaherigoldan: a couple things-
13:21jmaher1) make sure you are retriggering the right job, they are all &#39;throbber&#39; (t), but the specific job is in the perfherder details after clicking on the (t)
13:22jmaher2) some devices might be offline
14:30igoldanjmaher: are the Talos tests taking longer when used with the --geckoProfile flag?
14:31igoldanjmaher: initially, I thought there was a blocking bug
14:38jmaherigoldan: it should run longer, but each test will do 1 or 2 cycles, not 20 or 25 cycles
14:39igoldanjmaher: thanks
14:41igoldanjmaher: in 20 minutes we will have the weekly, right?
14:42jmaherigoldan: yes
14:42jmaherigoldan: a couple folks will not be there
17:10aswanis there anything that documents how the &quot;remote-nytimes&quot; test is measured?
17:10aswani&#39;m finding nothing at
17:11jmaheraswan: it is a smaller version of tp5o
17:11jmaheraswan: smaller so it runs faster on our limited physical phones
17:11jmaherit only runs on android
17:11aswanokay thanks
17:12jmaheraswan: apologies for lack of documentation- let me file a bug for that at least
17:12aswanjmaher: no worries, thanks for the quick answer
17:13aswanwhile we&#39;re here though, do you happen to know the specific timing? ie does the actual measurement start right after browser startup?
17:29jmaheraswan: sorry, paged away; as for specific timing, we launch the browser and it takes a second or two to get started, then the pages start loading and waiting for mozafterpaint+onload events and it reloads/repeats
17:41aswanjmaher: okay thanks!
22:07John-Galtmstange: Every time I try to run talos with the profiler now, I get a crash in GC code because previousGC is just uninitialized data in js::Nursery::renderProfileJSON
22:07mstangeJohn-Galt: florian filed a bug on that today
22:08* John-Galt will just hack around it for now
22:15florianmstange: is there a profiler API to easily print a stack of both C++ and JS from C++ code, for debugging purpose?
22:24mstangeflorian: apparently not; I thought there was one, but either it&#39;s been removed or I was imagining things
22:25mstangeflorian: if you want to write one (which uses profiler_get_backtrace()), I&#39;d happily r+ it
23:55sfinkbleh. Even after fixing that startup profiling crash that I introduced with the minorgc reporting stuff, profiling startup is extremely crashy.
18 Jul 2017
No messages
Last message: 8 days and 15 hours ago