Perf dashboard link: https://chromeperf.appspot.com/report?sid=ebe9a41281fb1870459fd3af339cb76973bf14821431439fb0fa2bd21bf7675d&rev=430152

Here are a few conclusions I've drawn about the source of the noise paired with the evidence that supports that conclusion:

** The noise is in fact noise, and isn't due to real changes in Chrome's power usage **

We can guess this because the ToT build and the ref build move together: https://chromeperf.appspot.com/report?sid=2cdc1b2d6faeaa2857dd97d5b6e15a0217044fe4d904d94d22989add8bf10bf8&rev=430152. If this weren't true, we'd only expect the ToT build to move.

** The source of the noise is long-lasting **

We can guess this because the noise doesn't go away between user stories. Instead, the power noise is highly correlated across user stories:

https://chromeperf.appspot.com/report?sid=e75a7347ee5a2efeb6c86c129e5e64a170c61b94dbf6f699812df16aadb5e817

We can also see this just by opening up a trace: power usage seems to be higher across the entire trace (min, average, and max), not just for a specific period of time during the trace.

Here are a few things I don't understand:

** The noise seems to impact the battor.trivial_pages benchmark much more than the battor.power_cases benchmark has significantly lower noise and moves between 12W and 13W. Note that in the following graph, there's a real regression (see regression.png). However, there's still noise, even without the large spike. See: https://chromeperf.appspot.com/report?sid=241f0e5afff23e0f4596001516843c318d77e48b992f35263e1ff4f28d212ebc

Our current hypothesis for this is that it's related to variable CPU temperatures, although I don't think this explains why it affects battor.trivial_pages so much more than battor.power_cases.
 

Deleted: regression.png 226 KB

	regression.png 226 KB View Download