"Flatten The Pick" Linux Patches Progress For Better cgroup Scheduling While Linux Gaming
A month ago I wrote about Linux scheduler work to help boost gaming performance on old "potato" hardware with Intel engineer Peter Zijlstra noting that Linux cgroup scheduling has continued to be "a pain in the arse." This work continues advancing with a third iteration of these "flatten the pick" patches being posted.
On Friday, which coincidentally was the 22nd birthday of Phoronix, a new version of these scheduling patches were posted to the Linux kernel mailing list that can make a noticeable impact on cgroup scheduling performance during Linux gaming.
The patches add a "cgroup_mode" knob and implements new policies for addressing the hierarchy level weight mismatch. ANd it builds upon that base for creating a flat/single run-queue scheduler where the cgroup hierarchy is expressed through dynamic weight management.
With Friday's updated patches, there are a number of fixes in place, the code has been re-based against the latest upstream state, cgroup_mode tasks is introduced, and the default cgroup_mode is changed. Peter Zijlstra's latest performance numbers for running a game on old Intel Sandy Bridge era hardware with a Radeon RX 580 Polaris graphics card while loading up with additional scheduling tasks:
Peter observed a huge improvement to the minimum frame rate and a healthy bump to the average FPS. The average and maximum frame times were also much lower.
The v3 patches are out for review on the Linux kernel mailing list to "flatten the pick" and will hopefully make it to the mainline kernel in the not too distant future.
On Friday, which coincidentally was the 22nd birthday of Phoronix, a new version of these scheduling patches were posted to the Linux kernel mailing list that can make a noticeable impact on cgroup scheduling performance during Linux gaming.
The patches add a "cgroup_mode" knob and implements new policies for addressing the hierarchy level weight mismatch. ANd it builds upon that base for creating a flat/single run-queue scheduler where the cgroup hierarchy is expressed through dynamic weight management.
With Friday's updated patches, there are a number of fixes in place, the code has been re-based against the latest upstream state, cgroup_mode tasks is introduced, and the default cgroup_mode is changed. Peter Zijlstra's latest performance numbers for running a game on old Intel Sandy Bridge era hardware with a Radeon RX 580 Polaris graphics card while loading up with additional scheduling tasks:
Peter observed a huge improvement to the minimum frame rate and a healthy bump to the average FPS. The average and maximum frame times were also much lower.
The v3 patches are out for review on the Linux kernel mailing list to "flatten the pick" and will hopefully make it to the mainline kernel in the not too distant future.