eACGM

eACGM: An eBPF-based Automated Comprehensive Governance and Monitoring framework for AI/ML systems.

English | 中文

---

:star: [News] Our work has been accepted by IEEE/ACM IWQoS 2025 (CCF-B)!

[arXiv]

---

eACGM provides zero-intrusive, low-overhead, full-stack observability for both hardware (GPU, NCCL) and software (CUDA, Python, PyTorch) layers in modern AI/ML workloads.

Features

• [x] Event tracing for CUDA Runtime based on eBPF
• [x] Event tracing for NCCL GPU communication library based on eBPF
• [x] Function call tracing for Python virtual machine based on eBPF
• [x] Operator tracing for PyTorch based on eBPF
• [x] Process-level GPU information monitoring based on libnvml
• [x] Global GPU information monitoring based on libnvml
• [x] Automatic eBPF program generation
• [x] Comprehensive analysis of all traced events and operators
• [x] Flexible integration for multi-level tracing (CUDA, NCCL, PyTorch, Python, GPU)
• [x] Visualization-ready data output for monitoring platforms

Visualization

To visualize monitoring data, deploy Grafana and MySQL using Docker. Access the Grafana dashboard at http://127.0.0.1:3000.

cd grafana/
sh ./launch.sh

Start the monitoring service with:

./service.sh

Stop the monitoring service with:

./stop.sh

Case Demonstration

The demo folder provides example programs to showcase the capabilities of eACGM:

• pytorch_example.py: Multi-node, multi-GPU PyTorch training demo
• sampler_cuda.py: Trace CUDA Runtime events using eBPF
• sampler_nccl.py: Trace NCCL GPU communication events using eBPF
• sampler_torch.py: Trace PyTorch operator events using eBPF
• sampler_python.py: Trace Python VM function calls using eBPF
• sampler_gpu.py: Monitor global GPU information using libnvml
• sampler_nccl.py: Monitor process-level GPU information using libnvml
• sampler_eacg.py: Combined monitoring of all supported sources
• webui.py: Automatically visualize captured data in Grafana

Citation

If you find this project helpful, please consider citing our IWQoS 2025 paper:

@misc{xu2025eacgmnoninstrumentedperformancetracing,
      title={eACGM: Non-instrumented Performance Tracing and Anomaly Detection towards Machine Learning Systems}, 
      author={Ruilin Xu and Zongxuan Xie and Pengfei Chen},
      year={2025},
      eprint={2506.02007},
      archivePrefix={arXiv},
      primaryClass={cs.DC},
      url={https://arxiv.org/abs/2506.02007}, 
}