1134591 - Consider sampling for per-compartment CPU monitoring

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine, defect)

Description

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

Bug 674779 introduces per-compartment CPU monitoring. The current approach is exact, but might be expensive in terms of CPU.

An alternative would be to adopt sampling, e.g. every n milliseconds, check `getrusage` off the main thread and attach the cost to the current compartment. A naive implementation, however, would probably cause Firefox to wakeup way too often, which would make it quite bad for power usage.

Comment 1

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

Sketching out a possible solution to adopt sampling without killing battery use. Essentially, the idea is that the timer thread should be able to go to sleep shortly after realizing that the main thread is inactive.

Atomic variables:
* bool mainThreadHeartBeat   // `true` if the main thread has processed at least one event recently
* bool isTimerThreadSleeping // `true` when the timer thread is waiting for the main thread
* resourceUsage // the result of `getrusage`


Main thread (in XPConnect's OnProcessNextEvent())
1. Set `mainThreadHeartBeat` to `true`;
2. If `isTimerThreadSleeping` is `true`
  2.1 set `isTimerThreadSleeping` to `false`;
  2.2 wake up timer thread;
3. Process events as usual.


Timer thread:
1. Sleep until awakened by the main thread;
2. Set `isTimerThreadSleeping` to `false`;
3. Wait for n milliseconds;
4. If `mainThreadHeartBeat` is `false`;
  4.1 set `isTimerThreadSleeping` to `true`;
  4.2 go to 1;
5. Call `getrusage` and put the result in `resourceUsage`;
6. Set `mainThreadHeartBeat` to `false`;
7. Goto 3.

Comment 2

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

Notes: the timer thread should probably not go to sleep immediately if `mainThreadHeartBeat` is `false`, but rather allow several missed heart beats, in case the main thread is just busy.

The main thread may also set `mainThreadHeartBeat` to `true` both when entering and leaving the event.

Comment 3

[Bill McCloskey [inactive unless it's an emergency] (:billm)]

You might want to look at the xpconnect watchdog thread.
http://mxr.mozilla.org/mozilla-central/source/js/xpconnect/src/XPCJSRuntime.cpp#980

It's responsible for triggering the slow script dialog. It seems like this functionality could be added to it without too much trouble though.

Comment 4

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

Good idea.

Comment 5

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

This doesn't seem necessary anymore.