Power Management and Energy-awareness

The Linux Plumbers 2017 Power Management and Energy-awareness track focuses on topics related to power management frameworks, task scheduling in relation to power/energy optimization, and platform power management mechanisms. The goal is to facilitate cross framework and cross platform discussions that can help improve power and energy-awareness in Linux.

Topic proposals may relate to (but not restricted to):

  • Frequency/performance scaling (cpufreq).
  • CPU idle states (cpuidle).
  • Thermal management and scheduling under thermal constraints.
  • ACPI power management.
  • Device Tree representation of platform power management features.
  • Energy-aware scheduling decisions (minimize energy consumption).
  • Power domains management.
  • Measurement techniques and tools.

Key Attendees

  • Rafael J. Wysocki
  • Morten Rasmussen
  • Kevin Hilman
  • Peter Zijlstra
  • Vincent Guittot
  • Lorenzo Pieralisi
  • Sudeep Holla
  • Srinivas Pandruvada
  • Len Brown
  • Ulf Hansson
  • Dietmar Eggemann
  • Juri Lelli
  • Patrick Bellasi
  • Viresh Kumar
  • Lina Iyer
  • Manoj Rao

Proposed Topics (Tentative)

Please add topic proposals here:

  • Linux Power Management Frameworks Status Report (Rafael J. Wysocki rafael@kernel.org): Overview of the current status of high-level PM frameworks in the Linux kernel.
  • Workloads for evaluation CPU PM modifications (Rafael J. Wysocki rafael@kernel.org): Discussion on which workloads to use (or what is recommended etc) for the evaluation of CPU PM changes in the kernel.
  • Energy-aware scheduling in user space (Len Brown lenb@kernel.org): Is it viable or can the kernel do much better? When is it better to use a “core packing” policy?
    • There are userspace implementations of EAS in the wild. I'm happy to talk about the one we wrote. (Michael Turquette mturquette@baylibre.com)
  • Identification of power-hungry system components (Srinivas Pandruvada srinivas.pandruvada@linux.intel.com): In thermally constrained situations it would be better to throttle the system components that draw more power instead of throttling everything equally. The problem is how to identify them reliably.
  • Resume from hibernation in a boot loader (Rafael J. Wysocki rafael@kernel.org): Is it viable at all? How much time can we save anyway?
  • schedutil usage/aggregation of utilization signals (Juri Lelli juri.lelli@arm.com): How do we aggregate utilization signals coming from different scheduling classes? Do we need an additional rt_rq/rt_se signal?
  • Adding a CPPC/HWP interface to cpufreq (Juri Lelli juri.lelli@arm.com): How to modify cpufreq to allow the setting of minimum guaranteed performance and desired performance level?
  • Atomic cpufreq set operation using platform firmware (Sudeep Holla sudeep.holla@arm.com): How to achieve or at-least make operations in OSPM atomic for CPU frequency scaling using platform firmware interface ?
  • Exploring aperf/mperf feature like on ARM systems (Sudeep Holla sudeep.holla@arm.com): With newer ARM platforms implementing some related counters, can we make use of them until interface is completely standardized ?
  • Scheduling under CPU capacity constraints (Morten Rasmussen morten.rasmussen@arm.com): The potential difference in max and delivered CPU performance (capacity) is growing (thermal and peak current management). Can we keep ignoring it, or should load-balancing factor the constraints in?
  • Per-Task capacity clamping (Patrick Bellasi patrick.bellasi@arm.com): Status update on the proposal to have a per-task API to based bias OPP selection and, in the future, task placement at wake-up time.
  • Improving the estimation of tasks utilization (Patrick Bellasi patrick.bellasi@arm.com): The dynamic nature of PELT makes it not completely suitable to describe the bandwidth requirements of long sleeping tasks and a prototype for “decay clamping” has been evaluated to be not suitable to tackle this issue. Let's review the status for a different proposal to add a filter on top of PELT.
  • An update on Window Assisted Load Tracking (Vikram Mulukutla markivx@codeaurora.org): There were a number of discussions on WALT and PELT in LPC 2016. We will present data on a recent Snapdragon 835 platform comparing PELT and WALT with the proposed PELT improvements. This will include a comparison of both task placement as well as CPU frequency guidance. We believe this discussion will be valuable to the continued effort to improve PELT, as well as the discussion of WALT's upstream-implementation potential, given its continued inclusion in the android-common kernels.
  • Next steps for PELT (Vincent Guittot vincent.guittot@linaro.org): Present and discuss the next steps for improving PELT.
  • Enable the runtime PM centric path for ACPI (Ulf Hansson ulf.hansson@linaro.org) Discuss about the benefits of using the runtime PM centric path for system sleep and in particular how it impacts the ACPI PM domain.
  • Next steps on CPU cluster idling (Ulf Hansson ulf.hansson@linaro.org) Unfortunately the work has moved slowly forward, however let's talk briefly on how far we have come and what to expect next.
  • Update on the generic PM domain (Ulf Hansson ulf.hansson@linaro.org) Highlight recent changes for genpd and an update on current ongoing developments.
  • The quest for faster Linux suspend/resume (Zhang Rui rui.zhang@intel.com) With test result from analyze_suspend tool, share some of the key issues fixed, and those still outstanding, especially some typical failures, which are likely to be repeated.
  • The current status of hibernation (Yu Chen yu.c.chen@intel.com) Collect/summarize the bugs information reported against hibernation, which are either fixed, or do not have proper solution, or not root caused yet, may also be related to other components such as memory subsystem, interrupt subsystem, cpu hotplug etc, which are more likely to be triggered with the increasing of cpu number.
  • Performance-centric “energy” model (Michael Turquette mturquette@baylibre.com) How to support heterogenous multicore systems that are not energy-sensitive, but are performance sensitive? Can the energy cost model be used?


Discussion notes

Will be added here when available.


Runners: Rafael J. Wysocki rafael@kernel.org, Morten Rasmussen morten.rasmussen@arm.com, Kevin Hilman khilman@baylibre.com

2017/power_management_and_energy-awareness.txt · Last modified: 2017/09/10 06:14 by
Except where otherwise noted, content on this wiki is licensed under the following license:CC Attribution-Noncommercial-Share Alike 3.0 Unported
Recent changes RSS feed Donate Powered by PHP Valid XHTML 1.0 Valid CSS Driven by DokuWiki