Processor Power States: C-States and P-States

• Software plays an important role in overall platform energy efficiency.
• Ill-behaving software thwarts hardware's power management benefits.
• Computing platforms combine hardware, software and technologies to run software.
• Data centers are concerned with electricity costs for servers and cooling.

Processor Power States

• CPUs have C-states and P-states for energy management.
• C-states define the degree to which the processor is sleeping
• P-states are performance states that define frequency levels

C-States

• In the C0 state the CPU is active
• The CPU can take the opportunity to rest or 'sleep'
• C-states are often referred to as sleep states
• Deeper sleep states save more energy, even for short periods (e.g., 100 µs).

P-States

• P-states define the frequency at which the processor runs.
• Intel processors use SpeedStep, AMD processors use PowerNow! or Cool’n’Quiet, and VIA processors use PowerSaver.
• P0 is maximum power and frequency.
• P1 is less than P0, with voltage or frequency scaled down.
• Pn is the lowest rated voltage and/or frequency.
• P-states save energy according to the equation: P = CV^2 f
• Reducing frequency leads to reduced power, assuming constant capacitance and voltage.

Energy Saving Software Techniques

• Software behavior influences the effectiveness of energy-saving features.
• Well-behaved software allows energy-saving features to work; poorly behaved software inhibits them.
• Techniques for energy-efficient players include read-ahead buffering and hardware accelerators for video decode.

Software Energy Efficiency Techniques

• Computational efficiency focuses on getting work done quickly with minimal energy consumption.
• Data efficiency reduces energy costs by minimizing data movement.
• Context awareness involves sensing the environment and reacting to changes.
• Idle efficiency minimizes power consumption when the system is running but not actively executing workloads.

Computational Efficiency

• Completing tasks quickly and returning to idle saves energy.
• Approaches include efficient algorithms, multithreading, vectorization, and uArch tuning.

Efficient Algorithms

• The choice of algorithms and data structures impacts an application's performance.
• Carefully consider architecture, design, algorithms, and data structures for better performance.

Multithreading

• Multithreading software improves performance and energy efficiency on modern multi-core platforms.
• Methodologies and libraries like OpenMP, OpenCL, and Thread Building Blocks are available.

Vectorization

• Better computational efficiency is achieved by vectorizing code using SIMD for instruction-level data parallelism.

Data Efficiency

• Data efficiency reduces energy costs by minimizing data movement.
• Effective data efficiency is achieved by using following methods
• Software algorithms that minimize data movement
• Memory hierarchies
• Application software that efficiently uses cache memories

Managing Disk I/O

• Consider HDD characteristics (rotational speed, seek time, latency, transfer rate) to optimize power and performance.
• Experimentation provides results and recommendations for energy usage of HDDs.
• Tests include:
  • Impact of block size on sequential reads
  • Effect of buffering during multimedia playback
  • Impact of file fragmentation
  • Impact of native command queuing on random reads
  • Disk I/O in multithreaded code

Impact of Block Size on Sequential Reads

• Hypothesis: Reading large volumes of sequential data in larger chunks uses less power.
• Reading large files in larger blocks leads to lower CPU utilization and energy consumption.
• Block sizes of 8 KB or greater can improve performance.

Impact of Native Command Queuing (NCQ) on Random Reads

• Hypothesis: Asynchronous I/O with NCQ improves performance and saves energy.
• Using NCQ reduces total time by ~15% with a similar reduction in energy.
• Use asynchronous I/O to take advantage of NCQ by Queuing requests and use events to determine complete.

Buffering During Multimedia Playback

• Hypothesis: Reading ahead and caching media content saves energy.
• Reading data in small chunks keeps the hard disk active and consumes more power.
• Utilize a buffering strategy in multimedia playback to minimize disk reads.

Impact of Fragmentation

• Hypothesis: Reading a fragmented file uses more energy.
• Fragmented files take longer to read.
• Avoid by pre-allocating large files and defragment volumes periodically.

Disk I/O in Multithreaded Code

• Hypothesis: Performance and energy costs of multithreading is reduced through data sharing coordination.
• Queuing I/O calls and utilizing NCQ provides disk I/O performance
• When multiple threads cause disk thrashing consolidate into read thread

Pre-Fetching and Caching

• Buffering reduces DVD power consumption by 70% and overall platform power consumption by 10%.
• Minimize DVD drive use by reducing spin-up, spin-downs, and read accesses.
• Let the operating system manage CPU frequency for P-state adjustments.

Context Awareness

• Sense the environment and react to changes.
• Requires sensors and state change events.
• Examples:
  • Low battery warning
  • Dimming display on DC power
  • Adjusting brightness to ambient light

Awareness of Power Source

• Important for applications to be aware of continuous power source (AC) or operating on battery (DC).
• Applications should register for event notifications.
• Important events switch from AC to DC power and reaching battery threshold

Platform Power Policies

• Windows provides power policies like High performance, Balanced, and Power saver'
• Adjust application behavior based on the user's current power policy.
• Change application behavior in response to a change in power policy.
• Change the power policy to suit the application behavior if necessary.

Other Context-Aware Behaviors

• Consider LAN cards and radios.
• Networking increases energy consumption.
• System Event Notification Service (SENS) can help mobile application issues.
• SENS API provides function calls for checking network connection and pinging addresses.

Idle Efficiency

• Idle Power: power consumed when the system is running in ACPI SO state with software applications/services running but not executing workloads.
• Aim to minimize background activity.
• Challenge: lower the idle floor by improve application idle efficiency

Deep C-State Residency

• Keep the platform in deeper C-states.
• Residency in the deepest C-state (C6/C7) should be more than 90%.
• Minimize C-state transitions.
• Frequent C-state transition behavior impacts power consumption.
• Do not split a task between processes/threads unless parallel execution is possible to reduce C-state transitions

OS Timer Resolution

• Default system-wide timer resolution in Windows is 15.6 ms.
• Timer interval decreased to less than 10 ms reduces system battery life by 25%.
• Transitions from low-power states incur energy cost.
• Use PowerCfg /energy to see if an application changed the platform timer resolution.
• Set resolution that meets specific platform demands

Background Activity

• Frequent periodic background activity increases power consumption.
• Impacts processor and chipset power.
• Long-running events prevent system idling.
• Windows 7 improvements to minimize frequent idle activity include:
  • Elimination of TCP distributed program call (DPC) timer on every system timer interrupt.
  • Reduction in frequency of USB driver maintenance timers
  • Intelligent timer tick distribution
  • Timer coalescing

Methodologies for Energy Conservation and Potential Benefits

• (See table)
• Includes techniques, and descriptions, and their benefits

Evaluating and Measuring Software Impact to Platform Power

• DAQ tools require an instrumented platform for precise power measurement.

Fluke NetDAQ (Networked Data Acquisition Unit)

• Measures platform power consumption while running applications.
• Has multiple programmable channels (10–100).
• Target PC has a special motherboard with built-in sense resistors wired and soldered.

Software Tools

Windows 7 PowerCfg

• Command line tool for system power management
• Detects common energy efficiency problems when system is idle.
• Output an HTML report of problems

PowerInformer

• Tool developed by Intel to provide basic power-relevant statistics to a developer.
• Measures and optimizes for Battery and power status

Energy Checker

• Measures software energy efficiency and write energy-aware software.
• Core API consists of five functions: open, reopen, read, write, and close a counter.
• Using the API exposes metrics of 'useful work' done by an application

Sustainable Software Development

• Software's indirect impact is real, including higher energy consumption and increasing e-waste.
• Software also creates social issues, such as lack of accessibility support.
• "Green" refers to environmental aspects; sustainability encompasses environmental, economic, and social dimensions.