George Wright (:gw280) (needinfo me!) Assignee
	Description • 8 years ago

This is our current data for GPU_PROCESS_LAUNCH_TIME_MS:

https://telemetry.mozilla.org/new-pipeline/dist.html#!cumulative=0&end_date=2016-11-14&keys=__none__!__none__!__none__&max_channel_version=nightly%252F53&measure=GPU_PROCESS_LAUNCH_TIME_MS&min_channel_version=null&product=Firefox&sanitize=1&sort_keys=submissions&start_date=2016-11-14&table=0&trim=1&use_submission_date=0

We're going to want to have some more insight into that highest bucket. Upon consulting people who know more about this than I do, it seems our best option is to add a GPU_PROCESS_LAUNCH_TIME_MS_LONG that specifically breaks down the last bucket.

	George Wright (:gw280) (needinfo me!) Assignee
	Comment 2 • 8 years ago

I'm not sure this needs to be an eternal probe, so I've set it to expire in fx 55 here. We can always renew it later.

	George Wright (:gw280) (needinfo me!) Assignee
	Comment 3 • 8 years ago

It might also be worth putting more telemetry markers in to break down the startup time - ie, how long it takes to actually do the process launch vs. how long to initialise whatever needs to be initialised.

	David Anderson [:dvander] - inactive, e-mail if emergency
	Comment 4 • 8 years ago

Comment on attachment 8811238 [details] [diff] [review]
0001-Bug-1317796-Add-a-GPU_PROCESS_LAUNCH_TIME_LONG_MS-pr.patch

Review of attachment 8811238 [details] [diff] [review]:
-----------------------------------------------------------------

Will this automatically exclude samples < 1000ms? If not, would be good to do that in GPUChild.

	George Wright (:gw280) (needinfo me!) Assignee
	Comment 5 • 8 years ago

I believe so. My understanding is that the "low: 1000" field in the histogram definition should cover that.

	Benjamin Smedberg
	Comment 6 • 8 years ago

It will not exclude them entirely: they will be included in the "low" bucket of the histogram, just as any measurements greater than 64000 would be in the "high" bucket. Which seems like the right thing to me.

	Benjamin Smedberg
	Comment 7 • 8 years ago

Why is this better than modifying the current exponential histogram with another 10 buckets and increasing it's max to 64000?

	George Wright (:gw280) (needinfo me!) Assignee
	Comment 8 • 8 years ago

I was told that can't be done? Chris?

	Chris H-C :chutten
	Comment 9 • 8 years ago

If I don't miss my guess, what bsmedberg is suggesting is the "Rebucket and rename" approach (see GC_REASON_2) as opposed to the "Add a second one for the long pauses" (see FX_TAB_SWITCH_SPINNER_VISIBLE_LONG_MS)

To my knowledge, neither is better than the other in implementation, ease-of-use, bandwidth cost, or analysis. (Since we've done both in similar cases)

	George Wright (:gw280) (needinfo me!) Assignee
	Comment 10 • 8 years ago

I would suspect then that this would be better, as we should be able to get the GPU process down to < 1000ms launch times for the vast majority of launches, and so the existing one should be useful long-term, and we can use this second one for now to try and get more data on how bad the situation is.

	David Anderson [:dvander] - inactive, e-mail if emergency
	Comment 11 • 8 years ago

Well, it's possible we can't do anything even if we learn that stuff is slow :)

But either way we should be able to turn one of the histograms off at that point.

	George Wright (:gw280) (needinfo me!) Assignee
	Comment 12 • 8 years ago

Would you prefer the rebucket and rename then? I think both are equally good for our needs.

	Benjamin Smedberg
	Comment 13 • 8 years ago

Comment on attachment 8816312 [details] [diff] [review]
0001-Bug-1317796-Rebucket-and-rename-GPU_PROCESS_LAUNCH_T.patch

Yeah I slightly prefer one histogram if it solves your problems equally well.

	Pulsebot
	Comment 14 • 8 years ago

Pushed by gwright@mozilla.com:
https://hg.mozilla.org/integration/mozilla-inbound/rev/3ea1f28bd0f3
Rebucket and rename GPU_PROCESS_LAUNCH_TIME_MS to cater for launch times up to 64000ms r=dvander data-review=bsmedberg

	Carsten Book [:Tomcat]
	Comment 15 • 8 years ago
bugherder

https://hg.mozilla.org/mozilla-central/rev/3ea1f28bd0f3