Oh, and as for how I expect they'll be implemented:
Time to interaction sounds straightforward to me - on a
per-script-thread basis, track the start of the last 10s window that did
not contain a task that took longer than 50ms to execute. Whenever at
least 10s has gone by since the start of this window, scan the set of
contained documents that have not yet had a timestamp recorded. If any
of them have dispatched the DOMContentLoaded event, mark their TTI
timestamp as max(DOMContentLoaded time, start of 10s window). If a task
takes longer than 50ms, reset the start of the window to the current
timestamp.
The painting ones are interesting, because the spec text is roughly
equivalent to the creation of display list items for the things we care
about. However, since that happens in parallel with returning control to
the script thread, it's not totally clear where to stop measuring. An
easy place to start is to send a message to the profiler directly from
the layout thread after notable display list items are created. However,
the spec is concerned about when control returns to the script thread
event loop, and I have not been able to come up with a good mechanism to
report that timestamp while allowing the layout thread to continue
creating display lists in parallel.
Cheers,
Josh
On 2017-03-31 4:18 PM, Josh Matthews wrote:
Back in September, Till shared an update [1] from dev-platform about the
Progressive Web Metrics that the Blink team is championing. I've been
looking into how that effort is going in Firefox and Blink, and whether
we should be adopting some of them.
Overview:
We should implement the Time to First Paint, Time to First Contentful
Paint, and Time to Interactive measurements, focusing on integrating
them with fitzgen's graphical profiler tool. The specifications make use
of the PerformanceObserver web API which we don't implement yet; if
we're keen on using these measurements to compare testcases against
other browsers, we should rectify that. It will not make it easier to
compare measurements for arbitrary web content, however.
Long form:
In brief, some of the measurements have pretty clear definitions and
would be worth integrating into Servo in Q2. Others are still being
hashed out and wouldn't clearly provide value to us at the moment.
Time to First Paint and Time to First Contentful Paint [2] are
conceptually straightforward. Respectively, they measure the number of
milliseconds before the browser paints anything that is not the default
white background, and the time before the browser paints any text,
image, non-white canvas, or SVG.
Firefox has already implemented Time to First Paint [3]. I think we'll
get the most value out of this measurement by focusing on the
interaction with our graphical profiler timeline tool, rather than the
integration with the new PerformanceObserver API [4].
Time to Interactive takes a different path. It attempts to measure the
amount of time before a page is ready for interaction, and has low
enough latency for meaningful interaction to occur. It tracks 10 second
windows in which no script thread task takes longer than 50ms to
complete. If a task takes longer, then the window is reset. To determine
readiness, the official write-up [5] uses the Time to First Meaningful
Paint metric (not discussed here), but there have also been experiments
using the DOMContentLoaded event for this purpose (ie. the page has
finished parsing). I propose implementing an initial version of this
metric in Servo using DOMContentLoaded as the signal.
Together, these metrics will give us information on the graphical
timeline showing:
* how long after a page was requested before the first visual change
occurred
* how long after that before the first nontrivial visual content appeared
* how long after the page was requested before it was in a state that
could respond to user input with low latency
If there are no objections, I intend to make implementing these part of
our Q2 goals.
Cheers,
Josh
[1]
https://groups.google.com/d/msg/mozilla.dev.platform/Ot43qpeWyos/x20WdRnxBAAJ
[2] https://github.com/WICG/paint-timing#definition
[3] https://bugzilla.mozilla.org/show_bug.cgi?id=1307242
[4] https://w3c.github.io/performance-timeline/#dom-performanceobserver
[5] https://github.com/tdresser/time-to-interactive
_______________________________________________
dev-servo mailing list
dev-servo@lists.mozilla.org
https://lists.mozilla.org/listinfo/dev-servo
