Hi,
On 30/01/17 15:52, Viresh Kumar wrote:
On 26-01-17, 14:40, Juri Lelli wrote:
I managed to run jankbench and youtube with and without this set on top of a schedutil backport for Pixel phones. I uploaded results here:
https://drive.google.com/open?id=0B0gETIMiqtYIOHhmUDJyNXpzQkE
Not sure everybody is familiar with what reported, so please don't hesitate to ask for any sort of clarification, but I couldn't find any evident difference between having or not having the set in. This is a very narrow set of workloads, though. So, are there any other Android workloads/benchmark/platforms positively affected by the changes?
Hi Juri,
I tried Hackbench, recentfling and galleryfling. The first two didn't gave anything special, but I do see improvements with galleryfling.
I tested with workload automation and the results are attached in this email.
Thanks for sharing. Any idea why galleryfling in particular is seeing benefits? I'm asking because it might be hard to defend the changes upstream without a clear understanding of why they are needed (and the standard hackbench scores don't seem to help us here :().
Best,
- Juri
Most noticeable is the results for 90th and 95th percentiles:
With remote wakeups Without remote wakeups 23/44 28/57
FWIW, I am using images provided by Steve (around 300 MBs). I am uploading them to google drive for Vincent. Lemme know if you also need a link.
-- viresh