1166889 - Deactivate addon monitoring when the system is too busy

Status: RESOLVED DUPLICATE of bug 1200172

(Toolkit :: Performance Monitoring, defect)

Description

[David Teller [:Yoric] - still alive but not very active]

When the system is too busy (in particular when it's swapping), the add-on monitor is going to blame add-ons although they are not necessarily the culprits. We should deactivate addon monitoring (and possibly jank monitoring) when the system is busy due to other processes.

Comment 1

[David Teller [:Yoric] - still alive but not very active]

Any idea on how we could get the information under Windows, Aaron? Under MacOS, André? On other OSes, anyone?

Comment 2

[David Teller [:Yoric] - still alive but not very active]

For Linux, see http://phoxis.org/2013/09/05/finding-overall-and-per-core-cpu-utilization/

Comment 3

[David Teller [:Yoric] - still alive but not very active]

Also asking Paul, as this may be a problem also encountered in WebRTC.

Comment 4

[Paul Adenot (:padenot)]

We have some sort of a CPU monitoring thing for WebRTC [0], that is used to back-off some costly processing bits. 

[0]: https://dxr.mozilla.org/mozilla-central/source/dom/media/systemservices/LoadMonitor.h

Comment 5

[David Teller [:Yoric] - still alive but not very active]

Any chance you could move this to toolkit?

Comment 6

[Paul Adenot (:padenot)]

I haven't written the code, it was gcp and pkerr iirc, I just happen to know it exists.

gcp, pkerr, any objection to moving the load monitor somewhere other people can use it? This would help Yoric to implement a more reliable add-on monitoring code to detect add-ons that are slowing down Firefox.

Comment 7

[André Reinald]

(In reply to David Rajchenbach-Teller [:Yoric] (use "needinfo") from comment #1)
> Any idea on how we could get the information under Windows, Aaron? Under
> MacOS, André? On other OSes, anyone?

For Mac, either use host_processor_info or host_statistics, they are defined in mach/... includes.
Useful info can be found here:
https://stackoverflow.com/questions/20471920/how-to-get-total-cpu-idle-time-in-objective-c-c-on-os-x

But if we already have working code to do that, it would make more sense to use it.

Comment 8

[Gian-Carlo Pascutto [:gcp]]

(In reply to Paul Adenot (:padenot) from comment #6)
> gcp, pkerr, any objection to moving the load monitor somewhere other people
> can use it? This would help Yoric to implement a more reliable add-on
> monitoring code to detect add-ons that are slowing down Firefox.

No, seems like a fine idea. It was always meant as a generic service, there just weren't any consumers outside the media code yet.

Comment 9

[David Teller [:Yoric] - still alive but not very active]

> No, seems like a fine idea. It was always meant as a generic service, there
> just weren't any consumers outside the media code yet.

Filed as bug 1167175.

Comment 10

[Paul Kerr [:pkerr]]

(In reply to Paul Adenot (:padenot) from comment #6)
> I haven't written the code, it was gcp and pkerr iirc, I just happen to know
> it exists.
> 
> gcp, pkerr, any objection to moving the load monitor somewhere other people
> can use it? This would help Yoric to implement a more reliable add-on
> monitoring code to detect add-ons that are slowing down Firefox.

It is designed to notify multiple interested parties about system load. Moving it to toolkit should not impact our use for WebRTC. The code already contains OS specific mechanisms (some mentioned above) for load monitoring. When I landed the Windows and OSX bits, I confirmed that the reports matched the load monitoring tools supplied with each OS.

Comment 11

[(No longer employed by Mozilla) Aaron Klotz]

(In reply to David Rajchenbach-Teller [:Yoric] (use "needinfo") from comment #1)
> Any idea on how we could get the information under Windows, Aaron? Under
> MacOS, André? On other OSes, anyone?

Windows does expose performance counter data that could be used for this. See https://msdn.microsoft.com/en-us/library/windows/desktop/aa371903%28v=vs.85%29.aspx for API info. You could use the "Pages/sec" counter described at https://technet.microsoft.com/en-us/library/cc778082%28v=ws.10%29.aspx. The only problem with that counter is that the threshold for what constitutes "excessive" paging is relative to the system itself. One suggestion from https://msdn.microsoft.com/en-us/library/aa746253%28v=bts.10%29.aspx is to "observe whether processor activity drops significantly as paging increases."

Comment 12

[Paul Kerr [:pkerr]]

(In reply to Aaron Klotz [:aklotz] (please use needinfo) from comment #11)
> (In reply to David Rajchenbach-Teller [:Yoric] (use "needinfo") from comment
> #1)
> > Any idea on how we could get the information under Windows, Aaron? Under
> > MacOS, André? On other OSes, anyone?
> 
> Windows does expose performance counter data that could be used for this.
> See
> https://msdn.microsoft.com/en-us/library/windows/desktop/aa371903%28v=vs.
> 85%29.aspx for API info. You could use the "Pages/sec" counter described at
> https://technet.microsoft.com/en-us/library/cc778082%28v=ws.10%29.aspx. The
> only problem with that counter is that the threshold for what constitutes
> "excessive" paging is relative to the system itself. One suggestion from
> https://msdn.microsoft.com/en-us/library/aa746253%28v=bts.10%29.aspx is to
> "observe whether processor activity drops significantly as paging increases."

There is code in LoadMonitor.cpp that contains OS specific mechanisms for determining per CPU and total system processing load.
https://dxr.mozilla.org/mozilla-central/source/dom/media/systemservices/LoadMonitor.cpp#485
https://dxr.mozilla.org/mozilla-central/source/dom/media/systemservices/LoadMonitor.cpp#379

Paging is not something that was measured directly for the WebRTC monitor.

Comment 13

[David Teller [:Yoric] - still alive but not very active]

Done differently in bug 1200172.