Simulating telemetry interactive boost

[Dave Rodgman]

Hi,

I've been looking at the CPU scheduler and noticed that on some ChromeOS platforms,  the scheduler sets CPU frequency to maximum for a short period in response to any user input (i.e. CPU frequency will stay high during scrolling).

This doesn't happen when simulating user inputs via Telemetry, as the kernel does not see any input events. The result is that Telemetry performance does not accurately measure what the user experiences.

The kernel patch was introduced in April 2015: https://chromium-review.googlesource.com/c/257075/

I've written a patch for Telemetry which simulates the kernel performance boost behaviour (on the CrOS platform backend) via writing to sysfs. On CrOS platforms which implement the interactivity boost, this improves top_25_smooth results consistent with the normal user experience. (On unaffected platforms, behaviour is unaffected).

Should I file a bug for this and upload a change for review?

Regards

Dave

[Ned]

How do you plan to make this change? Is this a configuration to ChromeOS bots or a CL to Telemetry?

--
You received this message because you are subscribed to the Google Groups "telemetry" group.
To unsubscribe from this group and stop receiving emails from it, send an email to telemetry+...@chromium.org.

[Dave Rodgman]

It's a CL to Telemetry.

Dave

[Ned]

We will be having a project next quarter about overhauling how Telemetry does these config tuning methods in a more systematic way. With 5 different platforms to support, we do need a clear API for adding "do these things on this platform for that benchmark to throttle CPU/GPU,..." type of operation. 

If this issue is urgent to you, I would suggest you figure a way to change the CPU frequency from chromeOS bots without touching Telemetry.

[Dave Rodgman]

Thanks Ned, that's useful to know. Potentially the approach I've taken could be a stepping stone towards a more general API for CPU/GPU frequency control - certainly I think it could help demonstrate that this sort of thing can improve the quality of Telemetry's benchmark results. It also helps bridge the gap until this project can be landed.

On some platforms I have seen a large (57%) reduction in dropped frames in top_25_smooth with this patch, so it is clear that it significantly impacts performance numbers.

I've pushed the patch here: 

https://codereview.chromium.org/2876843002

It is work in progress at this time - EAS-based platforms such as kevin (Samsung Chromebook Plus) are not yet covered - but it would be great to get some feedback.

thanks

Dave