PowerTrack

Measure energy consumption of Python code — CPU · GPU · RAPL · Time-based estimates

PowerTrack provides a unified API to profile power usage with multiple backends—from simple time-based estimates to hardware Intel RAPL counters and NVIDIA GPU power sampling. Use it as a decorator or function call to measure joules, watts, and execution time of any callable.

---

Table of Contents

• Features
• Installation
• Quick Start
• API Reference
• Backends
• Usage Examples
• Measurement Output
• Contributing

---

Features

• Multiple backends: Time-based, CPU-weighted, Intel RAPL, and NVIDIA GPU
• Auto-detection: Automatically uses the best available backend on your system
• Concurrent measurement: Run multiple backends simultaneously and aggregate results
• Flexible API: Use as a function call or decorator
• Zero-config fallback: Works out of the box with the time-based backend (no hardware required)

---

Installation

From source

Place the powertrack package in your Python path. From the directory containing powertrack:

export PYTHONPATH="${PYTHONPATH}:$(pwd)"
python -c "from powertrack import measure_power; print('OK')"

Core (minimal)

The core package works with the time-based backend only—no extra dependencies required.

Optional backends

Backend	Package	Command
CPU-weighted	psutil	pip install psutil
Intel RAPL	pyRAPL	pip install pyRAPL
NVIDIA GPU	pynvml	pip install pynvml

Install all optional backends:

pip install powertrack psutil pyRAPL pynvml

---

Quick Start

Using the decorator

from powertrack import measure_power

@measure_power(backend="time_based")
def compute_sum(n):
    return sum(i**2 for i in range(n))

result = compute_sum(1_000_000)
print(f"Energy: {result.energy_j:.2f} J")
print(f"Time:   {result.time_s:.3f} s")
print(f"Power:  {result.avg_power_w:.2f} W")
print(f"Output: {result.result}")

Using the function API

from powertrack import measure_callable

def heavy_work(x):
    return sum(i**2 for i in range(x))

m = measure_callable(heavy_work, 1_000_000, backend="time_based", return_result=True)
print(f"Energy: {m.energy_j:.2f} J, Result: {m.result}")

Auto backend selection

from powertrack import measure_power

@measure_power(backend="auto")  # Uses best available: RAPL → GPU → CPU-weighted → time-based
def my_function():
    # your code here
    pass

---

API Reference

measure_callable(func, *args, backend="auto", backend_opts=None, runs=1, warmup=0, return_result=True, **kwargs) -> Measurement

Measure energy consumption of a callable.

Parameter	Type	Default	Description
func	Callable	—	The function or callable to measure
*args	—	—	Positional arguments passed to func
backend	str | List | type	"auto"	Backend name(s), class(es), or instance(s). See Backend selection.
backend_opts	Dict	None	Options passed to backend constructors. See Backend options.
runs	int	1	Number of measured runs (results are averaged)
warmup	int	0	Number of warmup runs (not measured, used to stabilize caches/CPU)
return_result	bool	True	If True, include the function's return value in Measurement.result
**kwargs	—	—	Keyword arguments passed to func

Returns: A Measurement instance with aggregated energy, time, and power data.

Example:

m = measure_callable(my_func, arg1, arg2, kwarg=3, backend="cpu_weighted", runs=3)

---

measure(func, *args, **kwargs) -> Measurement

Alias for measure_callable. Same signature and behavior.

---

measure_power(backend="auto", backend_opts=None, runs=1, warmup=0, return_result=True)

Decorator that measures a function's energy consumption.

Parameter	Type	Default	Description
backend	str	"auto"	Backend to use
backend_opts	Dict	None	Options for the backend
runs	int	1	Number of measured runs
warmup	int	0	Number of warmup runs
return_result	bool	True	Include return value in Measurement.result

Example:

@measure_power(backend="cpu_weighted", backend_opts={"tdp_w": 45}, runs=3, warmup=1)
def train_model():
    # ...
    return model

---

Measurement

Dataclass holding measurement results.

Attribute	Type	Description
time_s	float	Average wall-clock time in seconds
energy_j	float	Average energy in joules
avg_power_w	float	Average power in watts (energy / time)
backend	str	Always "combined" when using measure_callable
details	Dict | None	Per-run and per-backend breakdown
result	Any	Return value of the measured function (if return_result=True)

Methods:

• to_dict() -> Dict — Convert to a dictionary (excludes result). Useful for JSON serialization.

Example:

m = measure_callable(foo, backend="time_based")
print(m.to_dict())
# {'time_s': 0.5, 'energy_j': 7.5, 'avg_power_w': 15.0, 'backend': 'combined', 'details': {...}}

---

Backends

Backend selection

The backend parameter accepts:

Value	Behavior
"auto"	Try, in order: rapl → nvidia_gpu → cpu_weighted → time_based
"time_based"	Simple duration × constant power (no sensors)
"cpu_weighted"	CPU-time–weighted estimate (requires psutil)
"rapl"	Intel RAPL hardware counters (requires pyRAPL, Intel CPU)
"nvidia_gpu"	NVIDIA GPU power sampling (requires pynvml, NVIDIA GPU)
["a", "b"]	List of backends; all run concurrently, energies summed
Class/instance	Custom backend class or instance

Backend options

Pass options via backend_opts:

Global (all backends):

measure_callable(func, backend="time_based", backend_opts={"avg_power_w": 25})

Per-backend:

measure_callable(func, backend=["time_based", "cpu_weighted"],
                 backend_opts={
                     "time_based": {"avg_power_w": 20},
                     "cpu_weighted": {"tdp_w": 45}
                 })

---

Time-based backend (time_based)

No dependencies. Estimates energy as duration × avg_power_w.

Option	Type	Default	Description
avg_power_w	float	15.0	Assumed average system power (W)

Use case: Baseline, CI, or environments without sensors.

from powertrack.backends import TimeBackend

@measure_power(backend="time_based", backend_opts={"avg_power_w": 25.0})
def my_func():
    pass

---

CPU-weighted backend (cpu_weighted)

Requires: psutil

Estimates power as TDP × (cpu_time / wall_time). Calibrate tdp_w for your CPU.

Option	Type	Default	Description
tdp_w	float	28.0	CPU TDP in watts (thermal design power)

Use case: Better CPU-bound workload estimates than time-based.

@measure_power(backend="cpu_weighted", backend_opts={"tdp_w": 65})
def cpu_intensive():
    return sum(i**2 for i in range(10**7))

---

RAPL backend (rapl)

Requires: pyRAPL, Intel CPU with RAPL support (Linux)

Uses hardware energy counters for package (PKG) and DRAM.

Option	Type	Default	Description
devices	List[str]	None	e.g. ["PKG", "DRAM"] (pyRAPL.Device)
socket_ids	List[int]	None	Socket IDs to measure

Use case: Accurate CPU/DRAM energy on supported Intel systems.

@measure_power(backend="rapl", backend_opts={"devices": ["PKG", "DRAM"]})
def measured_code():
    pass

---

NVIDIA GPU backend (nvidia_gpu)

Requires: pynvml, NVIDIA GPU and drivers

Samples GPU power during execution via NVML.

Option	Type	Default	Description
index	int	0	GPU device index
sampling_interval	float	0.1	Sampling interval in seconds

Use case: GPU workload energy profiling.

@measure_power(backend="nvidia_gpu", backend_opts={"index": 0, "sampling_interval": 0.05})
def gpu_inference():
    # CUDA/GPU code
    pass

---

Usage Examples

Multiple runs and warmup

@measure_power(backend="time_based", runs=5, warmup=2)
def benchmark():
    return expensive_computation()

# 2 warmup runs (not measured) + 5 measured runs; results averaged
m = benchmark()

Multiple backends (concurrent)

m = measure_callable(
    my_func,
    backend=["time_based", "cpu_weighted"],
    backend_opts={
        "time_based": {"avg_power_w": 20},
        "cpu_weighted": {"tdp_w": 45}
    }
)
# Energies from both backends are summed
print(m.details["per_run"][0]["per_backend"])

Discard return value

@measure_power(backend="time_based", return_result=False)
def side_effect_only():
    write_to_disk()
    # No need to store return value

Custom backend instance

from powertrack import measure_callable
from powertrack.backends import TimeBackend

be = TimeBackend(avg_power_w=30.0)
m = measure_callable(my_func, backend=be)

---

Measurement Output

Structure

Measurement(
    time_s=0.123,           # Average wall time (s)
    energy_j=1.85,          # Average energy (J)
    avg_power_w=15.04,      # energy / time (W)
    backend="combined",
    details={
        "runs": 1,
        "active_backends": ["time_based"],
        "per_run": [
            {
                "wall_time_s": 0.123,
                "run_energy_j": 1.85,
                "per_backend": {
                    "time_based": {
                        "duration": 0.123,
                        "energy_j": 1.85,
                        "avg_power_w": 15.0
                    }
                }
            }
        ]
    },
    result=333332833333500  # Function return value
)

Backend-specific fields in per_backend

Backend	Extra keys in per_backend[name]
time_based	—
cpu_weighted	cpu_time_s
rapl	pkg_j, dram_j
nvidia_gpu	samples_count, samples

---

Project Structure

powertrack/
├── __init__.py          # Public API: measure, measure_callable, measure_power, Measurement
├── core.py              # measure_callable, backend resolution
├── decorators.py        # measure_power decorator
├── types.py             # Measurement dataclass
├── backends/
│   ├── __init__.py      # TimeBackend, CPUWeightedBackend
│   ├── time_based.py    # Time-based backend
│   ├── cpu_weighted.py  # CPU-weighted backend (psutil)
│   ├── rapl.py          # Intel RAPL backend (pyRAPL)
│   └── gpu.py           # NVIDIA GPU backend (pynvml)
└── README.md

---

Testing

Run the test suite:

pytest powertrack/test.py -v

From the project root with PYTHONPATH set:

PYTHONPATH=. pytest powertrack/test.py -v

---

Contributing

Contributions are welcome. Please open an issue or pull request on GitHub.

---

License

[Specify your license here]