mozilla :: #perf

11 Jul 2017
12:13mreiddthayer: nice analysis - if you're interested, you could add it to http://reports.telemetry.mozilla.org
13:24rwoodigoldan: thanks for the review
13:25igoldanrwood: you're welcome
15:08rwoodmconley: ping
15:08mconleyrwood: pong!
15:08rwoodmconley: hey, TIME_TO_NON_BLANK_PAINT (https://bugzilla.mozilla.org/show_bug.cgi?id=1377251) is not the same as MozAfterPaint with a non-empty rect, correct?
15:09mconleyrwood: I actually think it is
15:09rwoodhmm interesting
15:09rwoodso maybe I should use that, instead
15:10mconleyrwood: sounds like a plan - though you'll need to wait until a non-blank paint has occurred
15:10mconleyso I guess you could poll until the value is non-zero or something
15:11rwoodmconley: if it's a standard event handler I could just not remove it until it is non-zero
15:11RyanVMmconley, rwood: do we care about bug 1372261 enough to backport to 55? It grafts cleanly and could land a=test-only
15:11mconleyrwood: that sounds like a plan
15:12mconleyRyanVM: it'd be a harmless backport, yeah
15:12RyanVMk, I'll throw it in the queue
15:12rwoodcool!
17:42digitaraldmystor: "It takes some time after startup until BHR is running, and this test started running before we had finished initializing it" ... answers the question we had yesterday
17:42digitaraldso BHR skips most of startup
17:43mystordigitarald: Well, it's complicated
17:43mystordigitarald: BHR is initialized really early on in startup (we don't spin a single event off of the event loop before we start it), but on linux we then have to initialize lul
17:43mystorAnd initializing LUL is really really slow, so we do it off main thread on the BHR thread
17:44mystorAnd BHR can't do any work until LUL is initialized, so we don't detect any hangs which occur while that's happening. On windows and macOS it should be faster
17:50digitaraldmystor: ok, makes sense
21:33John-Galtmstange: Did something change to make profile collection a lot slower recently? When I try to collect profiles from local talos runs now, everything is an order of magnitude slower than it used to be. If I use sub-millisecond sampling intervals, the ts_paint test actually times out...
21:34mstangeJohn-Galt: I'm not aware of anything
21:34mstangeJohn-Galt: how recently?
21:34mstangewell, there have been a few changes to the platform profiler code
21:34John-Galtmstange: I don't know. I noticed it for the first time today, but I haven't done any profiling in the past few weeks.
21:34mstangeI see
21:34John-GaltI could try to bisect, I guess
21:34mstangethat would be extremely useful
21:35John-GaltWill do, then
22:33John-Galtmstange: Never mind. The problem went away when I switched to artifact builds. Apparently it was my own fault for adding -Og to my mozconfig and forgetting about it...
22:34mstangeJohn-Galt: oops
22:34mstangehmm
22:34mstangeso this made nm slower?
22:34mstangeoh no
22:34mstangeit made the whole build slower
22:34mstangeok
22:34John-GaltNo, it made everything slower, especially with the profiler enabled
12 Jul 2017
No messages
   
Last message: 13 days and 2 minutes ago