1. Chapter 9 Contributed by Alex Turek
Power Optimization
Chapter 9 Contributed by Alex Turek

2. Outline
System Design Considerations Power Optimization (hardware) Power Optimization (software) Constant vs. Dynamic Power in ES Advanced Configuration and Power Interface (ACPI) Constructing software for maximum power performance Summary

3. Computer System Design Considerations
Cost Performance Power Size
Every decision is a trade off
Generally, lowering power consumption has become more important than increasing performance
Initially developed for mobile platforms
Power optimization is important for all platform types, from mobile devices and embedded systems to multi-million dollar servers

4. Power Basics
Power can be optimized on both the hardware and software levels
Current computers consist of many MOSFETs (Metal Oxide Semiconductor field Effect Transistor)

5. Active Power Consumption
P = CV2f
C = capacitive load
V = voltage
f = Frequency
Voltage affects active power more so than the frequency or capacitive load of a circuit due to its exponential relation
Voltage and frequency are first order determinants of performance

6. Power Consumption in CMOS
Charging and discharging capacitors
Voltage and frequency are first order determinants of performance
Short circuit currents:
Short circuit between supply rails during switching
Leakage (static power consumption)
Leaking diodes and transistors
Minimizing leakage is extremely important for
low-power systems, which are not active most
of the time

7. Intel 3D Transistor
Intel's 3-D Tri-Gate transistors enable chips to operate at lower voltage with lower leakage The 22nm 3-D Tri-Gate transistors provide up to 37 percent performance increase at low voltage versus Intel's 32nm planar transistors the new transistors consume less than half the power when at the same performance as 2-D planar transistors on 32nm chips.

8. Intel 3D Transistor

9. Optimizing Power Using Software
Power Profiles of Embedded Systems:
Always on/continuous power
for devices that are required to provide near-peak performance, nearly all the time (displays, controllers, networking devices, etc)
Advanced Power Management such as Intel SpeedStep™
Policy and usage based frequency and voltage control
Initially designed for laptops: plugged in vs battery modes
Newer Intel CPUs such as the Core i7 series handle usage power management, where as the OS is used for policy based power management
Intel Core i7-3960X 3.3 GHz (3.9 Ghz Turbo) 6 core 12 thread

10. Reducing Power Consumption
Mobile computers/devices created the need for CPUs with better power consumption
Now, CPUs are using less of the overall system power budget, manufacturers are looking for ways to improve the power consumption of other components
The First Step to Optimizing System Power
Maximize the range of power consumption between low-power states and high power states
If the max/min states have negligible differences in power consumption, then power management will have no impact

11. Constant vs. Dynamic Power
Constant Power: an embedded computer consumes a constant, minimum amount of power when powered on, regardless of the level of system activity Other than by powering a system down completely or transitioning to a sleep state, software cannot influence a system’s constant power Dynamic Power: power consumption based on load If most of the components in the system support low-power modes, then dynamic power consumption can be significantly controlled through the careful use of resources

12. Idle – Peak Ratio
Ratio of Idle to Peak power consumption: estimate: 6/9 = 0.66;observed: 9/11 = 0.81 ~75% of total system power consumption is completely independent of software-level activity on the platform In comparison: Intel Core i7-3960X 3.3 GHz (3.9 Ghz Turbo) 6 core 12 thread Idle Power: 62 W;Peak Power: 210 W 62/210 = At 0% utilization, 29.5% of peak power consumed
FIGURE 9.1 Advertised and Measured Dynamic Power Range for CompuLab’s fit-PC2. The Advertised Power Range Covers 66% of Peak, whereas the Measured Value, When Accounting for Active I/O Devices, Covers 81% of the Observed Peak.

13. FIGURE Top: Power Consumption and Efficiency for a System With a Dynamic Power Range Down to 50% of Peak Performance. Power is Normalized to Peak Consumption. Efficiency is Calculated by Dividing Utilization by Normalized Power.
FIGURE Bottom: Power Consumption and Efficiency for a System With a Dynamic Power Range Down to 10% of Peak Performance.
Goal of Power Efficiency: high efficiency over entire operating range

14. ACPI
ACPI: (Advanced Configuration and Power Interface): platform specific industry standard that defines the roles, responsibilities, and specific mechanisms provided to achieve reliable power management and system configuration Developed through the cooperation of HP, Intel, Microsoft, Phoenix Technologies, and Toshiba Previous interfaces replaced by ACPI such as plug-and-play BIOS and Advanced Power Management (APM) were platform and operating system specific If most of the components in the system support low-power modes, then dynamic power consumption can be significantly controlled through the careful use of resources

15. ACPI continued…
The operating system is granted exclusive control over all system-level management tasks, including the boot sequence, device configuration, power management, and external event handling (such as thermal monitoring or power button presses)
ACPI Registers: well defined locations that can be read and
written to monitor and change the status of hardware
resources
ACPI BIOS: This firmware manages system boot and
transitions between sleep and active states.
ACPI Tables: describe the interfaces to the underlying
hardware and represent the system description.
Written in a domain specific language, ACPI Machine
Language (AML). The operating system's ACPI driver includes an interpreter for AML.
These tables are provided by the ACPI BIOS

16. Idle vs. Sleep
Two dominant forms of user-facing power management Sleep States: system appears to be powered off completely, while in fact some software may periodically execute to service external events, such as network packet arrivals. First introduced with laptop use in mind Idle Modes: take advantage of the variation in system utilization when the system is powered up Transitions between idle modes are made in milliseconds to reduce active power consumption, without the involvement or awareness of the user

17. ACPI System States
Global System States (Gx States) Describe the entire system Transitions between these states are typically observed by the user G0: Working G1: Sleeping (suspend or “instant boot” mode) G2 (S5): Soft Off G3: Mechanical Off

18. ACPI System States Continued
Sleep States (Sx States)
Sleep states within the global sleep state G1
S1: All system context is maintained by hardware upon entry and thus provides the lowest waking latency
S2: Operating system is responsible for storing and restoring CPU and cache hierarchy context. Upon entry to S2, CPU and cache state are lost
S3: Powers down more internal units than S2, but otherwise is similar
S4: Write all system states, including main memory, to nonvolatile storage
S5: Does not store system context.
Enables system to be powered on electronically (Wake on LAN)

19. ACPI System States Continued
Device Power States (Dx States)
Device class-specific approach to power management that allows devices of the same type to be treated in the same way
Device Classes: Audio, COM Port, Display, Input, Modem, Network, PC Card Controller, Storage
D0 (Fully On): fully powered up, fully operational state. Highest Power Consumption
D1: Initial sleep state for a device. Capable of waking itself or the entire system in response to an external event
D2: Incremental state beyond D1. Operates at a lower power and requires a greater waking latency than D1
D3hot: Saves a device-specific state so it can be awakened without a complete reboot
D3 (Off, or d3cold): Complete power down of the device. Device context not restorable

20. ACPI System States Continued
Processor Power States (Cx States)
ACPI allows processors to sleep while in the G0 working state. The Cx States only apply when the global state is G0.
C0: The processor executes instructions and operates normally
C1: Lowest transition latency CPU sleep state. Latency is so low that it is negligible and thus not a deciding factor in the operating system’s decision to transition to this state.
C2: Lower-power sleep state than C1. Operating system does consider the transition latency when deciding whether to transition to this state or another.
Transitions to both C1 and C2 are not apparent to software and do not alter system operation.
C3: Offers great power reduction at the cost of an increased transition latency. Processor caches maintain their state but do not emit cache coherence traffic
C4: Optional additional power states included in ACPI revision Even lower power consumption and greater transition latency. Vendor specific.

21. ACPI System States Continued
Processor Performance States (Px States)
Within processor power state C0, ACPI defines a range of performance states that are intended to enable a fully working system to vary its power consumption and performance by operating at different voltage and frequency levels. The operating system explicitly controls transitions between these states. For multiprocessor systems, each processor must support the same number and type of processor performance state.
P0: Maximum performance, maximum power consumption state.
P1: Next highest performing processor performance state and is expected to have second-greatest power consumption
Pn: ACPI allows for a maximum of 16 distinct performance states.

22. More on Intel SpeedStep Technology
Features performance and power management through voltage and frequency control, thermal monitoring, and thermal management features
Provides a central software control mechanism through which the processor can manage different operating points
In a FSB-based system, processors drive seven output voltage identification (VID) pins to facilitate automatic selection of processor voltages VCC from the motherboard voltage regulator
Each core in a multi-core has its own machine state register (MSR) to control the VID value. However, each core must work at the same voltage and frequency.

23. Software Construction for Maximum Power Performance
Linux PowerTOP Tool
Open-source power profiling tool for Linux that reports the occupation frequency of processor sleep and performance states
Race to Sleep
Modern systems transition between power states in response to load
Software should be organized to complete all available useful work in a continuous batch and then transition to an idle state and avoid unnecessary interruptions
Traditional software development tools emphasize code execution frequency.
Modern software dev tools must also take into account how often a program interferes with transitions to low-power CPU states

24. Evaluating Software and Systems with PowerTOP
PowerTOP can be used to measure and improve performance of code as it is written
The measurement harness is implemented in Python. To begin, the developer identifies the target code for measurement, referred to as the code-under-test (CUT)
Basic Operation of the Measurement Harness
Phase 1. Discover the maximum rate of execution of the CUT
Phase 2. Measure the power efficiency at the maximum execution rate
Phase 3. Measure the power efficiency as execution rate is scaled down from maximum
Two key parameters manipulated in Phase 3:
- total fraction of “sleep” time during the experimental observational period
- duration of an individual sleep interval

25. Concluding Remarks
System Design Considerations – “Everything’s a tradeoff”
Active Power Consumption – P = CV2f
Idle vs. Peak Ratio – decreasing constant power consumption is most favorable
Advanced Configuration and Power Interface
(ACPI) – Standard developed by the industry to provide universal power management
Software techniques for measuring and managing power consumption – race to sleep