Towards Power Management for FreeBSD

Transcription

1 Towards Power Management for FreeBSD Robin Randhawa FreeBSD Developer Summit Computer Laboratory University of Cambridge August 2015

2 Agenda An overview of Energy Aware Scheduling (EAS) for Linux Discussion on possibilities for FreeBSD

3 Motivations for EAS Hardware topologies are becoming more varied, accommodating different power/ performance budgets SMP, multi-cluster SMP, ARM big.little technology Per core/per cluster DVFS (Dynamic Voltage & Frequency Scaling) Linux power management frameworks are uncoordinated and hard to tune for different topologies cpufreq vs cpuidle vs scheduler The Task Scheduler is best placed to orchestrate power-performance control

4 Conventional Scheduling Scheduling policy decides task placement Affects performance and energy consumption Capacity Waking task Utilization Mainline Linux policy is work preserving Only cares about maximizing throughput DVFS and idle-states controlled by independent policy governors. Designed for SMP, not energy-aware Max capacity Current capacity? CPU cluster 0 1

5 Energy Aware Scheduling Energy-Aware Scheduling (EAS) policy Pick CPU with sufficient spare capacity and smallest energy impact Waking task Requirements Capacity? Utilization Tracking of task Platform energy model Supports all topologies SMP big.little Async DVFS Max capacity big Max capacity little Current capacity CPU cluster little big

6 Scale invariant load No invariance MHz MHz Throughput t MHz Running Area doubled t Utilization doubled t MHz Throughput ~Same area t ~Same t 0 t t 0 t

7 Component energy model Tabular cost data for all power and frequency domains in the system V V V Voltage regulator G G G Power Gate Power Domain Cluster (L2, SCU, ) G G Cluster (L2, SCU, ) G G C Clock Source Core 0 Core 1 Core 2 Core 3 G Clock Gate G G G G C C

8 Energy model data P-States (frequencies) Power Compute capacity Performance score normalised to the highest P-state of the fastest CPU in the system (1024) Busy power Normalised power score (W) Big core C-States (Idle states) Idle power (normalised) Normalised power score (W) Little core Compute capacity (Performance)

9 Estimating the energy impact Waking task Power Max capacity big Capacity? Utilization Placing task on CPU3: No P-state changes. CPU 2, 3 Big core Max capacity little Current capacity CPU 0, 1 Little core Placing task on cpu1: P-state change for CPU0 and CPU1. CPU cluster little big Compute capacity (Performance)

10 Idle state awareness Integration of cpuidle with the scheduler improves task placement on idle CPUs Waking task Scheduler picks CPU in shallowest idlestate (cheapest from a power and performance standpoint) Capacity? Utilization Max capacity Current capacity CPU cluster 0 1

11 Conventional DVFS Sampling based governors are slow to respond and hard to tune Sampling too fast: OPP* changes for small spikes cpu cpufreq governor sampling 0% 30% 100% React! Sampling too slow: Sudden burst of might not get the necessary OPP change in time. cpu 20% 20% 20% Time *OPP: Operating performance point (Voltage, frequency) tuple Average too low, no response. Time

12 Scheduler driven DVFS With scheduler task tracking DVFS can be notified immediately when CPU changes Improved responsiveness. Waking task Capacity? Utilization Cpu Now Next sample Max capacity big Max capacity little Capacity Current capacity Time CPU cluster little big

13 Centralised tunability Current: A set of governor-specific tunables. Goal: Single tunable to bias the energy/ performance trade-off. Prototypes Global boost tunable: /proc/sys/kernel/sched_cfs_boost Task group (cgroup) based tuning: /sys/fs/cgroup/stune/<group>/ schedtune.boost cpu Capacity Performance margin Time

14 EAS related tools/utilities rt-app synthetic workload generator for Linux ARM Workload Automation runs Android/ChromeOS tests Kernelshark trace analysis Improved analysis tools (experimental) TRAPpy: Trace Analysis and Plotting in Python BART: Behaviour Analysis and Regression Toolkit

15 Where do we go from here? (Very) early thoughts on incremental steps for FreeBSD Settle on a suitable platform Good documentation for powerperf machinery Measurement capability DAQs etc Has to be CPU subsystem specific Abstract task load scale Modular load calc runqueue residency time micro-arch hints I/O subsystem driven constraint specification for CPU power management Device tie-in Abstract capacity scale for processing elements PE topology expression Energy model expression SMP big.little Modular architecture Justin Hibbit s clock control Thermal sense driven constraint specification for task scheduling and CPU power management Energy model driven Thermal Advanced Incremental steps Basic Simple DVFS support Simple cpuidle support Modular policy Driven from the scheduler Scheduler aware Task wakeup Centralised tunability At one end: Energy efficient operation at best possible throughput At the other end: Outright performance at the intentional expense of energy Scheduling visualisation Tooling DVFS and Idle visualisation Energy model build flows

16 End