★ Session Summary ★
Scene setting (Morten)
Workshop scheduled session wise update by attendees that ran/participated-in those sessions (Mike, Daniel, Nico, Morten, Dietmar).
Key outcomes:
- Understanding of what the maintainer opinion was on current work
- Understanding on next steps for each track (EAS core, sched <-> PM interop, tools)
Tools supporting EAS
- Brief overview since there is already a detailed session earlier in the week: LCU14-206)
- Idlestat, rtapp (Amit and Vincent)
- Current status
- Feature roadmap
---------------------------------------------------
★ Resources ★
Zerista: http://lcu14.zerista.com/event/member/137769
Google Event: https://plus.google.com/u/0/events/cssb4s1l4mrhjvga49jphg808o8
Presentation: http://www.slideshare.net/linaroorg/lcu14-402-kernel-summitupdatefinal-39262924
Video: https://www.youtube.com/watch?v=UhmfeCqQrxU&list=UUIVqQKxCyQLJS6xvSmfndLA
Etherpad: http://pad.linaro.org/p/lcu14-402
---------------------------------------------------
★ Event Details ★
Linaro Connect USA - #LCU14
September 15-19th, 2014
Hyatt Regency San Francisco Airport
---------------------------------------------------
http://www.linaro.org
http://connect.linaro.org
2. Workshop details
Full-day workshop held at Kernel Summit 2014, August 19th
2014.
Workshop goal:
Coordinate plans for developing/improving energy-awareness in the
kernel scheduler and coordination with power management
frameworks.
Attendees from the entire community including Intel, Google,
IBM, Linaro, QualComm, ARM.
Summaries:
2
http://lwn.net/Articles/609561/
http://lwn.net/Articles/609969/ (detailed)
3. Workshop Agenda
Topics Owner
Benchmarking and testing energy improvements Daniel Lezcano (Linaro),
3
Mike Turquette (Linaro)
Task/cpu utilization tracking, compute capacity
representation and load/utilization tracking
improvements
Morten Rasmussen
(ARM)
Load tracking continued and scheduler/cpufreq
interaction
Mike Turquette (Linaro),
Preeti U Murthy (IBM)
Improving idle state selection Daniel Lezcano (Linaro),
Preeti U Murthy (IBM)
Platform performance/energy data/model in scheduler or
user-space
Morten Rasmussen
(ARM)
Power management in virtualized environments Preeti U Murthy (IBM)
5. Energy-aware Scheduling Workshop
Tools and benchmarks:
5
Workload generation tool based on RT-app is under development by Linaro:
Generates test patterns described used JSON
Two workload descriptions available already: Android audio playback and
web broswing.
Ftrace P- & C-state analysis tool (idlestat) by Linaro:
Gathers power/energy related statistics
Plans for integrating platform power model support to estimate energy
consumption.
Outcome
Community feedback on tools
Extended and agree on suite of workload (JSON) descriptions so we can get
a common set of benchmarks.
Limit scope to cpu energy consumption. There are no current proposals for
how to integrate memory, GPU, or other peripheral energy consumption.
6. Energy-aware Scheduling Workshop
Load and utilization representation in the scheduler:
6
Agreement on including a utilization metric in addition to the current
load tracking. The scheduler documentation needs to be updated with
terminology and concepts.
Agreed that we need to make per-entity load tracking (and utilization)
scale invariant, i.e. factor in frequency scaling.
Next step (and somewhat agreed) is to let the scheduler drive
frequency scaling (event driven) to get rid of existing sampling-based
governors (long term). The current cpufreq governors will continue to
exist while scheduler driven frequency scaling is being developed.
Energy-awareness support must be optional (possible to compile
out/static_key) for platforms that don't care or have hardware magic
that can't cooperate with the OS.
7. Energy-aware Scheduling Workshop
Energy model driven scheduling proposal:
7
Techniques for energy-aware scheduling are not beneficial for all platforms.
To avoid getting too many tunable heuristics that are nearly impossible for
non-scheduler developers to tune, it has been proposed to have a simple in-kernel
platform specific energy model.
The model will be used to estimate the implications of scheduling decisions,
not provide accurate absolute energy estimates.
The model is basically a set of P- and C-state tables, which should be
reasonable to ask platform vendors to provide. Alternatively, it should be
possible to extract on platforms where energy/power measurement is
possible.
Concept RFC patch set for ARM TC2 is out. Needs to be tested on more
platforms.
Model data to be provided by arch/* code. May come from DT, ACPI, or
driver.
Agreement/no violent objections on continuing work on this idea.
8. Energy-aware Scheduling Workshop
Idle-state awareness:
Lack of scheduler idle-state awareness needs to be addressed.
Important for both PowerPC and ARM.
Should be addressed as part of the energy model proposal.
Work on better prediction of wakeup events discussed.
Power policies under virtualization:
8
Communicating power management constraints from guest to host.
Not desirable. Power management should be controlled only by the
host.