On Friday 17 May 2019 at 09:43:46 (+0100), Quentin Perret wrote:
Hi,
+Patrick on CC
Now done ...
On Friday 17 May 2019 at 08:01:19 (+0000), Zhifei Yang (Arm Technology China) wrote:
(Post the questions here on behalf of Yun Hsiang)
Hi all,
I’d like to ask discuss some questions about uclamp in linux-power.githttp://www.linux-arm.org/git?p=linux-power.git;a=summary.
Code is based on the following version. linux-power.githttp://www.linux-arm.org/git?p=linux-power.git;a=summary tag: topic-util_clamp-20190412
- If a task clamped utilization is higher than a cpu capacity, should EAS bypass that cpu?
kernel/sched/fair.c static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) { … /* Skip CPUs that will be overutilized. */ util = cpu_util_next(cpu, p, cpu); <= Should this util use task clamped utilization? cpu_cap = capacity_of(cpu); if (cpu_cap * 1024 < util * capacity_margin) continue; … }
In short, yes, EAS should take this into account. The upstream uclamp implementation only targets frequency selection for now, as a first step. Patches for biasing task placement will come later, in a second step. But, yes, this is correct, and we have plans to implement this feature.
On a side note, if these questions come in the context of a product development, I just wanted to mention that our current recommendation is to use android kernels rather than the integration branch as a base. Up to android-4.19 schedtune is still the recommended option for products. One of the reasons is indeed that uclamp is not yet feature-complete.
- Will uclamp add a kernel-space API to set per-task utilization clamping?
Drivers/kernel module can use this API for precise performance/power tuning.
ISTR this has already been suggested but I can't recall the detailed outcome of the discussion. Patrick ?
Thank you for the feedback, it is very appreciated.
Regards, Quentin _______________________________________________ eas-dev mailing list eas-dev@lists.linaro.org https://lists.linaro.org/mailman/listinfo/eas-dev