Thanks Benjamin for kicking off this process understand how lower level metrics fit into the higher level quality tracking.
To better understand the long tail of slow process startup, 95th percentile is over 5s, the context of this metric could : 1) How often is a content process started during a user's session? This would define if we should track the long tail or median (probably both). 2) Probably related to Gabor's question: How does this metric fit into the breakdown of timings from interaction to tab loading? Is it the major part or are there other metrics we need to consider? 3) Do we know the circumstances that cause the long tail? /Harald On Tue, Feb 7, 2017 at 9:57 AM, Gabor Krizsanits <gkrizsan...@mozilla.com> wrote: > Thanks a lot for these numbers! Also, is there a way to get numbers for > the time from process launch to the point when we start loading the first > URL in the new content process? We run a lot of frame and process scripts > sooner than we would get there, and I'm afraid we would get even worse > numbers that way, and that will affect user experience. (I'm not sure how > to filter the related telemetry based on "if the new tab is starting up a > new content process or not") > > For a temporary workaround until we can speed up content process start up > and initialization time, we might want to use the pre-allocated process > manager for the e10s-multi case. Maybe even force to pick one from the > existing processes unless there is one already available. > > Gabor > > On Tue, Feb 7, 2017 at 6:36 PM, Benjamin Smedberg <benja...@smedbergs.us> > wrote: > >> https://telemetry.mozilla.org/new-pipeline/dist.html#!cumula >> tive=0&end_date=2017-02-06&keys=__none__!__none__!__none_ >> _&max_channel_version=aurora%252F53&measure=CONTENT_PROCESS >> _LAUNCH_TIME_MS&min_channel_version=null&product=Firefox& >> sanitize=1&sort_keys=submissions&start_date=2017-01-26& >> table=0&trim=1&use_submission_date=0 >> >> This shows the distribution of times to launch a content process from the >> time we initially ask for it to the time we get back the first >> initialization message through IPDL. So this covers actual process launch >> in the OS, XPCOM startup, and other bootstrap. >> >> At first glance, this appears worrying to me: almost 25% of content >> process >> startups take more than 1 second, and the median is >700ms. And this is on >> nightly/aurora, which users typically have faster computers. >> >> There's a lot of potential noise here: we don't know what else is going on >> on the computer (maybe it's near boot and there's still a lot of system >> churn). But this time definitely can have an impact on how quickly Firefox >> is ready to load pages and therefore the impression that users have of its >> total speed. >> >> Soliciting everyone's opinion, but Harald's in particular: is it important >> to dive into this in more detail soon (before Firefox 57)? >> >> This metric is currently exploratory, and so I need guidance about whether >> it's important to keep this metric around for e.g. a release-health >> dashboard or to prevent regressions. >> >> --BDS >> _______________________________________________ >> dev-platform mailing list >> dev-platform@lists.mozilla.org >> https://lists.mozilla.org/listinfo/dev-platform >> > > _______________________________________________ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
