Wednesday, 01-12-2010




IRC Nick

Amit Kucheria


Amit Arora


Vishwanath Sripathy


Yong Shen


Vincent Guittot


Amit Kachhap



New Action Items

  • Yong: Document how to bind a driver (cooling) to a thermal sensor and look at OMAP4 for examples of thermal drivers
  • AmitDaniel: Try cpufreq-bench and cyclic tests to exercise menu governor problems

  • Vincent: Talk to partners about what their plans for hotplug are

Old Action Items

  • Immediate:
    • ACTION: Srinivas: Check if Versatile Express has any thermal sensors: Pending
    • ACTION: Yong: Start a guide to thermal framework at DONE

    • ACTION: Srinivas: Enquire about status of talking to community about the causality tool for power profiles: Pending
    • ACTION: Amitk: Paraphrase our requirements for PM_QoS for dealing with application constraints vs. system constraints: Pending
    • ACTION: Vincent: Verify with Vishwa that cpuidle statistics only generated from common code (not platform-specific driver): DONE
    • ACTION: Vishwa to work with Amit Kucheria on re-targetting/re-prioritizing BLOCKED work items in cpuidle-latency-infrastructure blueprint: DONE
    • ACTION: Robin: Get back with details on how to split access into secure and non-secure context: PENDING
    • ACTION: amitdaniel: Discuss with Vincent regarding improvements to menu governor for performance: DONE
    • ACTION: amitdaniel: Add work item to investigate a simpler governor that cares only about wakeup latency w/o the prediction logic: PENDING
    • ACTION: amitdaniel: Add work item to investigate how the governor can mask certain c-states based on latency constraints: PENDING
    • ACTION: ARM: results of using PMU counters for correlation to PM: POSTPONED
    • ACTION: ARM to share internal instrumentation flow (BAB: we might also align with Linaro on workload discussions): POSTPONED
    • ACTION: Srinivas to provide details of where he believes userspace - kernel interaction is required. (low prio): INPROGRESS


  • Document thermal interfaces
  • Cpuidle statistics
    • Discussion about the mismatch between what is reported in cpuidle statistics and what idle states the cpu actually reaches
    • Actual idle transition happen in platform-specific cpuidle driver, hence that information not available to generic code that shows statistics
      • Do we need to enhance the statistics with this information?
    • cpuidle tracepoint work being done upstream
      • tracepoints show actual latency in platform-specific driver
      • so, if the governor's target state was OFF, but we only reached RETENTION, then the latency would be much less than expected
      • Indirect way of knowing that we didn't reach target state
  • Menu governor performance
    • ST-E did some work to remove data-prediction logic from governor to improve performance, hard to make it work for general case
    • Amit D. have problem coming up with standard usecases to test performance
  • Miscellaneous
    • Mini-sprint in Bangalore next week
    • Will talk to developers at partner companies for ideas on how to help them debug power problems better

WorkingGroups/PowerManagement/Meetings/2010-12-01 (last modified 2010-12-01 10:45:03)