Enerji İdarəetməsi

Enerji Problemi

graph TB
    Power[Power Consumption] --> Dynamic[Dynamic Power<br/>Switching activity]
    Power --> Static[Static Power<br/>Leakage]
    
    Dynamic --> Formula1[P = α × C × V² × f]
    Static --> Formula2[P = V × I_leakage]
    
    Formula1 --> Alpha[α: Activity factor]
    Formula1 --> Cap[C: Capacitance]
    Formula1 --> Volt[V: Voltage]
    Formula1 --> Freq[f: Frequency]

Dynamic Power:

P_dynamic = α × C × V² × f

α: Activity factor (0-1, işləyən transistorlar)
C: Capacitance (load)
V: Voltage
f: Frequency

V² → Voltage-ın kvadratik təsiri!

Static Power (Leakage):

P_static = V × I_leakage

Modern CPUs: Static power ~30-50% of total
Reason: Smaller transistors → more leakage

Total Power:

P_total = P_dynamic + P_static

Reducing voltage → Reduces both!

Dynamic Voltage and Frequency Scaling (DVFS)

DVFS Əsasları

graph LR
    High[High Performance<br/>High V, High f<br/>High Power] --> Medium[Medium<br/>Medium V, Medium f<br/>Medium Power]
    Medium --> Low[Low Performance<br/>Low V, Low f<br/>Low Power]
    
    Low -.Workload increases.-> High
    High -.Workload decreases.-> Low

Əsas prinsipler:

Frequency ↓ → Performance ↓ (linear)
Voltage ↓ → Power ↓ (quadratic!)

Example:
Freq: 3 GHz → 1.5 GHz (50% performance)
Voltage: 1.2V → 0.9V
Power: 1.2² → 0.9² = 1.44 → 0.81 (44% power savings!)

Voltage-Frequency Relationship

graph TB
    V1[Voltage 1.2V] --> F1[Max Freq: 3 GHz]
    V2[Voltage 1.0V] --> F2[Max Freq: 2 GHz]
    V3[Voltage 0.8V] --> F3[Max Freq: 1 GHz]
    
    Note1[Higher voltage → faster switching]
    Note2[Lower voltage → slower, but saves power]

Critical path delay:

Delay ∝ V / (V - V_threshold)²

Lower voltage → longer delay → lower max frequency

P-states (Performance States)

P-states - CPU frequency və voltage levels (işləyərkən).

graph TB
    P0[P0: Turbo Boost<br/>3.8 GHz, 1.35V<br/>Highest performance<br/>Highest power] --> P1[P1: Base<br/>3.0 GHz, 1.2V]
    P1 --> P2[P2<br/>2.0 GHz, 1.0V]
    P2 --> P3[P3<br/>1.5 GHz, 0.9V]
    P3 --> Pn[Pn: Lowest<br/>800 MHz, 0.8V<br/>Lowest performance<br/>Lowest power]

P-state Məlumatları

# Linux - P-state info
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_available_frequencies
# Output: 3800000 3000000 2000000 1500000 800000 (kHz)

cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_cur_freq
# Current frequency

cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
# powersave, performance, ondemand, conservative, schedutil

P-state Governors (Linux)

graph TB
    Governors[CPU Freq Governors] --> Performance[performance<br/>Always max frequency]
    Governors --> Powersave[powersave<br/>Always min frequency]
    Governors --> Ondemand[ondemand<br/>Dynamic based on load]
    Governors --> Conservative[conservative<br/>Gradual freq changes]
    Governors --> Schedutil[schedutil<br/>Scheduler-driven<br/>Modern default]

ondemand:

// Pseudo-code
if (cpu_load > 80%) {
    set_frequency(max_frequency);
} else {
    scale_frequency_proportionally(cpu_load);
}

schedutil:

// Integrated with scheduler
frequency = f(utilization, deadline, priorities);
// More responsive than ondemand

Intel P-state Driver

# Check driver
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_driver
# intel_pstate or acpi-cpufreq

# Intel P-state modes
cat /sys/devices/system/cpu/intel_pstate/status
# active, passive, off

# Turbo enable/disable
cat /sys/devices/system/cpu/intel_pstate/no_turbo
# 0 = turbo enabled, 1 = disabled

# Min/Max performance percentage
echo 50 > /sys/devices/system/cpu/intel_pstate/min_perf_pct
echo 100 > /sys/devices/system/cpu/intel_pstate/max_perf_pct

AMD P-state Driver

# AMD P-state (newer CPUs)
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_driver
# amd-pstate or acpi-cpufreq

# Similar interface to Intel

C-states (Idle States)

C-states - CPU idle (işləməyərkən) enerji qənaəti.

graph TB
    C0[C0: Running<br/>Active execution<br/>Full power] --> C1[C1: Halt<br/>Clock gated<br/>~1µs wake-up]
    C1 --> C2[C2: Stop Clock<br/>PLL running<br/>~10µs wake-up]
    C2 --> C3[C3: Sleep<br/>L1/L2 flushed<br/>~100µs wake-up]
    C3 --> C6[C6: Deep Sleep<br/>Core powered off<br/>~1ms wake-up]
    
    C6 -.Wake up.-> C0
    
    style C0 fill:#f66
    style C1 fill:#fc6
    style C2 fill:#ff6
    style C3 fill:#6f6
    style C6 fill:#6cf

C-state Levels

State	Name	Power	Wake latency	What's off/on
C0	Operating	100%	0	Everything on
C1	Halt	~90%	~1 µs	Clock gated, instant wake
C2	Stop-Clock	~70%	~10 µs	PLL on, caches on
C3	Sleep	~40%	~100 µs	L1/L2 flushed to L3
C6	Deep Sleep	~10%	~1 ms	Core power off, L3 on
C7	Deeper Sleep	~5%	~2 ms	LLC (L3) flushed
C8/C10	Package C-states	~1-2%	~5 ms	Entire package sleep

C-state Məlumatları

# List available C-states
ls /sys/devices/system/cpu/cpu0/cpuidle/
# state0  state1  state2  state3 ...

# C-state info
cat /sys/devices/system/cpu/cpu0/cpuidle/state3/name
# C6

cat /sys/devices/system/cpu/cpu0/cpuidle/state3/latency
# 200 (microseconds)

cat /sys/devices/system/cpu/cpu0/cpuidle/state3/power
# 0 (relative power consumption)

# Usage statistics
cat /sys/devices/system/cpu/cpu0/cpuidle/state3/time
# Total time spent in this state (microseconds)

cat /sys/devices/system/cpu/cpu0/cpuidle/state3/usage
# Number of times entered

Disable C-states (for latency-sensitive apps)

# Disable C-states deeper than C1
for state in /sys/devices/system/cpu/cpu*/cpuidle/state*/disable; do
    if [[ $(basename $(dirname $state)) != "state0" ]] && \
       [[ $(basename $(dirname $state)) != "state1" ]]; then
        echo 1 > $state
    fi
done

# Or via kernel parameter
# Add to GRUB: intel_idle.max_cstate=1

C-state Governor

# cpuidle governor
cat /sys/devices/system/cpu/cpuidle/current_governor
# menu, ladder, teo

# menu: Default, heuristic-based
# ladder: Simple, step-by-step
# teo: Timer events oriented (modern)

Turbo Boost / Turbo Core

Temporarily exceed base frequency.

graph TB
    Base[Base Frequency<br/>3.0 GHz<br/>All cores, sustained] --> Turbo[Turbo Boost<br/>Up to 4.5 GHz<br/>Few cores, temporary]
    
    Turbo --> Conditions{Conditions}
    
    Conditions --> Power[Power budget<br/>TDP not exceeded]
    Conditions --> Thermal[Temperature<br/>Below threshold]
    Conditions --> Current[Current<br/>Below limit]
    Conditions --> Cores[Core count<br/>Fewer active cores<br/>→ Higher boost]

Intel Turbo Boost

Intel Core i7-12700K:
- Base: 3.6 GHz (P-cores)
- Turbo: Up to 5.0 GHz (single core)
- All-core turbo: ~4.7 GHz

Depends on:
- Power (TDP: 125W base, 190W turbo)
- Thermal (< 100°C)
- Number of active cores

AMD Precision Boost

AMD Ryzen 9 7950X:
- Base: 4.5 GHz
- Boost: Up to 5.7 GHz

Precision Boost 2: More fine-grained control
Precision Boost Overdrive (PBO): User-configurable limits

Thermal Management

graph TB
    Temp[Temperature] --> T1[< 70°C<br/>Normal operation<br/>Full performance]
    T1 --> T2[70-90°C<br/>Thermal monitoring<br/>Turbo limited]
    T2 --> T3[90-100°C<br/>Throttling<br/>Reduce frequency]
    T3 --> T4[> 100°C<br/>Critical<br/>Emergency shutdown]
    
    style T1 fill:#6f6
    style T2 fill:#ff6
    style T3 fill:#f96
    style T4 fill:#f66

Thermal Throttling

T_junction (max): 100°C (typical for modern CPUs)

If temp > threshold:
1. Reduce frequency (P-state down)
2. Reduce voltage
3. Insert idle cycles (clock modulation)
4. Last resort: Shutdown (prevent damage)

Thermal Sensors

# Linux sensors
sensors
# Output:
# coretemp-isa-0000
# Core 0:       +45.0°C  (high = +80.0°C, crit = +100.0°C)
# Core 1:       +47.0°C

# Per-core temperature
cat /sys/class/hwmon/hwmon*/temp*_input
# Temperature in millidegrees (45000 = 45°C)

CPU Cooling Solutions

graph LR
    Air[Air Cooling<br/>Heatsink + Fan<br/>~50-80W TDP] --> Liquid[Liquid Cooling<br/>AIO / Custom loop<br/>~200-300W TDP]
    Liquid --> Exotic[Exotic<br/>Phase change, LN2<br/>Extreme OC]
    
    Air --> Cheap[Cheaper]
    Liquid --> Better[Better cooling]
    Exotic --> Expensive[Expensive, impractical]

Race-to-Idle

Philosophy: Finish work quickly, then sleep deeply.

sequenceDiagram
    participant Workload
    participant CPU
    participant Power
    
    Note over Workload: Work arrives
    Workload->>CPU: Task
    
    CPU->>CPU: Boost to max frequency
    Note over CPU: P0 state<br/>High power, fast
    
    CPU->>Workload: Task complete (fast!)
    
    CPU->>CPU: Deep sleep
    Note over CPU: C6 state<br/>Very low power
    
    Note over Power: Average power: LOW<br/>(short high + long low)

Race-to-Idle vs Constant Low

Scenario: Process 1 second of work

Option 1: Constant low frequency (1 GHz)
- Duration: 1 second @ 10W = 10 Joules
- Total: 10 Joules

Option 2: Race-to-idle (3 GHz → C6)
- Duration: 0.33 sec @ 30W + 0.67 sec @ 1W = 10 + 0.67 = 10.67 Joules
- Total: 10.67 Joules

BUT: Option 2 allows deeper sleep states longer
+ Lower latency (finishes faster)
+ Better for bursty workloads

Benefits

1. Lower latency - Work completes faster
2. Better sleep - More time in deep C-states
3. Power efficiency - Deep sleep saves more power
4. Responsiveness - UI feels snappier

When NOT to use Race-to-Idle

Sustained workloads (rendering, compilation):
- No idle time to sleep
- Constant high power
→ Better to use moderate frequency

Package C-states (PC-states)

Entire CPU package (all cores) idle.

graph TB
    PC0[PC0: Active<br/>At least one core active] --> PC2[PC2: Package idle<br/>All cores in C6<br/>LLC on]
    PC2 --> PC6[PC6: Deep<br/>LLC flushed<br/>Package power gated]
    PC6 --> PC8[PC8: Deeper<br/>Lower voltage]
    PC8 --> PC10[PC10: Deepest<br/>Minimal power<br/>Longest wake-up]
    
    PC10 -.Wake.-> PC0

Requirements for PC6:

• All cores in C6
• GPU idle (integrated graphics)
• I/O idle

# Check package C-state
cat /sys/kernel/debug/pmc_core/package_cstate_show
# PC2: 10%
# PC6: 80%
# PC10: 5%

RAPL (Running Average Power Limit)

Intel enerji monitoring və limiting.

graph TB
    RAPL[RAPL Domains] --> PKG[Package<br/>Entire CPU]
    RAPL --> PP0[Power Plane 0<br/>CPU cores]
    RAPL --> PP1[Power Plane 1<br/>GPU (integrated)]
    RAPL --> DRAM[DRAM<br/>Memory]
    
    PKG --> Limit[Power Limit<br/>TDP constraint]

RAPL Monitoring

# Install powercap utils
# Read package power
cat /sys/class/powercap/intel-rapl/intel-rapl:0/energy_uj
# Energy in microjoules

# Power limit
cat /sys/class/powercap/intel-rapl/intel-rapl:0/constraint_0_power_limit_uw
# Power limit in microwatts

# Set power limit (requires root)
echo 50000000 > /sys/class/powercap/intel-rapl/intel-rapl:0/constraint_0_power_limit_uw
# 50W limit

RAPL with perf

# Measure power consumption
perf stat -e power/energy-pkg/,power/energy-cores/,power/energy-ram/ ./program

# Output:
#   24.50 Joules power/energy-pkg/
#   20.10 Joules power/energy-cores/
#    4.20 Joules power/energy-ram/

Mobile Power Management

ARM big.LITTLE / DynamIQ

graph TB
    Scheduler[OS Scheduler] --> Light[Light tasks<br/>LITTLE cores<br/>Low power]
    Scheduler --> Heavy[Heavy tasks<br/>big cores<br/>High performance]
    
    Light --> Example1[Background sync<br/>Music playback<br/>UI idle]
    Heavy --> Example2[Gaming<br/>Video encoding<br/>Browsing]

Energy efficiency:

big core (A78): 1.0x performance, 1.0x power
LITTLE core (A55): 0.3x performance, 0.1x power

For light tasks: 3x more energy efficient!

Intel Hybrid (Alder Lake+)

graph TB
    ThreadDirector[Thread Director<br/>Hardware hints] --> P[P-cores<br/>Performance<br/>Golden Cove]
    ThreadDirector --> E[E-cores<br/>Efficiency<br/>Gracemont]
    
    P --> Foreground[Foreground apps<br/>Games<br/>Heavy workloads]
    E --> Background[Background tasks<br/>OS services<br/>Light workloads]

Power Profiling

powertop (Linux)

sudo powertop

# Tabs:
# - Overview: Power consumption estimate
# - Idle stats: C-state residency
# - Frequency stats: P-state usage
# - Device stats: Per-device power
# - Tunables: Optimization suggestions

# Generate HTML report
sudo powertop --html=report.html

# Auto-tune (enable all power savings)
sudo powertop --auto-tune

Intel Power Gadget

# macOS/Windows tool
# Shows:
# - Package power (W)
# - Frequency (GHz)
# - Temperature (°C)
# - Utilization (%)

Battery Life Optimization

# Check wakeups (frequent wakeups prevent deep sleep)
sudo powertop
# Look for processes with high wakeup count

# Reduce wakeups
# - Timer coalescing
# - Batch network I/O
# - Reduce polling

# Laptop mode (Linux)
echo 5 > /proc/sys/vm/laptop_mode
# Reduce disk writes, batch I/O

Performance vs Power Trade-offs

graph TB
    Axis[Performance vs Power] --> MaxPerf[Max Performance<br/>Servers, Desktops<br/>Always P0, C-states off]
    Axis --> Balanced[Balanced<br/>Laptops<br/>DVFS enabled, moderate C-states]
    Axis --> MaxPower[Max Power Savings<br/>Mobile, IoT<br/>Aggressive sleep, low freq]
    
    MaxPerf --> UseCase1[Minimize latency<br/>Predictable performance]
    Balanced --> UseCase2[Good performance<br/>Reasonable battery]
    MaxPower --> UseCase3[Long battery life<br/>Tolerate latency]

Configuration Examples

Max Performance (Latency-sensitive):

# Disable C-states > C1
intel_idle.max_cstate=1

# Performance governor
cpupower frequency-set -g performance

# Disable turbo (for consistency)
echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo

Balanced (Default):

# Default C-states
# schedutil governor (default)

# Turbo enabled
echo 0 > /sys/devices/system/cpu/intel_pstate/no_turbo

Max Power Savings (Battery):

# Enable all C-states

# Powersave governor
cpupower frequency-set -g powersave

# Reduce max frequency
echo 50 > /sys/devices/system/cpu/intel_pstate/max_perf_pct

# Enable powertop recommendations
powertop --auto-tune

Future Trends

1. Finer-grained Power Control

Modern: Per-core DVFS
Future: Per-cluster, per-instruction domain

2. AI-based Power Management

Machine learning predicts workload
→ Proactive frequency/C-state selection
→ Better than reactive governors

3. Heterogeneous Integration

CPU + GPU + NPU + DSP on same package
→ Route tasks to most efficient processor

4. Better Race-to-Idle

Faster wake-up from deep sleep
→ More aggressive deep sleep
→ Better average power

Best Practices

Development

1. Profile power consumption

   perf stat -e power/energy-pkg/ ./app

2. Reduce wakeups

   // BAD: Frequent polling
   while (true) {
       check_status();
       usleep(10000);  // Wake every 10ms
   }
   
   // GOOD: Event-driven
   poll(fds, nfds, -1);  // Sleep until event

3. Batch operations

   // Process multiple items per wake-up
   // Allows longer sleep periods

4. Use hardware acceleration

   // Hardware decoders use less power than software

System Administration

1. Choose right governor

   • Desktop: schedutil or ondemand
   • Server: performance (if latency critical)
   • Laptop: powersave or schedutil

2. Monitor thermals

   watch -n 1 sensors

3. Check power limits

   cat /sys/class/powercap/intel-rapl/intel-rapl:0/constraint_0_power_limit_uw

4. Verify C-state residency

   # Should spend most idle time in deep C-states
   sudo powertop

5. Tune for workload

   • Batch jobs: Max performance
   • Interactive: Balanced
   • Idle servers: Power savings

Əlaqəli Mövzular

• CPU Architecture: Frequency, voltage domains
• Modern Architectures: big.LITTLE, hybrid P/E cores
• Performance: Power-performance trade-offs
• Thermal Design: Cooling solutions
• Mobile Computing: Battery life optimization