gpulink 0.1.0

A simple tool for monitoring and displaying GPU stats

gpulink

A library for monitoring and displaying NVIDIA GPU stats. For this purpose gpulink provides classes for data recording and plotting, targeting different GPU stats (e.g. memory consumption). gpulink makes use of pynvml - a Python wrapper for the NVIDIA Management Library (NVML).

Feature	Status	Notes
MemoryRecorder	✅
MemoryPlotter	✅

During installation, gpulink also registers a command-line script accessible through the gpulink command.

Library usage

To integrate gpulink to a Python script, import gpulink and create an NVContext. This context manages the creation and destruction of the nvml session and provides several utility functions:

import gpulink as gpu

with gpu.NVContext() as ctx:

Every subsequent gpulink call (e.g. instantiation of a data recorder) must be made within this context. The context provides also additional information, i.e. about the numbers and names of compatible GPU devices:

    print(f"Available GPUs: {ctx.gpu_names}")

gpulink provides data recorders (e.g. MemoryRecorder) for capturing and storing GPU stats and plotters (e.g. MemoryPlotter) for rendering these:

    recorder = gpu.MemoryRecorder(ctx)
    mem_info = recorder.record(store=False)

    for _ in range(5):  # Record memory information over time
        recorder.record()

    records = recorder.get_records()
    for rec in records:
        print(f"Sampling rate: {rec.sampling_rate}")
        print(f"Recording duration: {rec.duration}")
        print(f"Number of records: {rec.len}")
        print(rec)

    graph = gpu.MemoryPlotter(records)
    graph.draw_graph()

Command-line usage

usage: gpu-link.py [-h] [-o OUTPUT]

GPU-Link memory tracing

optional arguments:
  -h, --help            show this help message and exit
  -o OUTPUT, --output OUTPUT
                        Path to memory graph plot.

Example

1. Run recording
   (a) Installed from source: python -m gpulink -o memory_consumption.png
   (b) Installed using pip: gpulink -o memory_consumption.png
2. Stop recording Ctrl+C
3. Plot is saved to memory_consumption.png.

Output

NVIDIA TITAN RTX[0] -> memory used: min=1729.236992[MB] / max=2315.79648[MB]