pPerf: Multi-Tenant DNN Inference Predictability Profiler

pPerf is a suite of tools for profiling, benchmarking, and analyzing multi-model DNN inference pipelines for autonomous driving. The toolkit supports single- and multi-model inference with LiDAR, image, and multi-modal (BEVFusion) models using ROS 2 for real-time and offline replay. It includes profiling tools for timing, GPU/CPU usage, and NVTX-based CUDA tracing. Data can be published from NuScenes or replayed from ROS 2 bags. Logs include timing, predictions, and resource usage in JSON/CSV. Post-processing tools provide kernel, layer, and system-level analysis, with scripts for running standard benchmarks.

Quick Start

1. Install
   Follow this to setup: INSTALL.md.

2. Prepare your data
   Download and extract the NuScenes dataset. Update paths in your config if needed. See DATA.md.

3. Launch the pipeline

   • Start the data publisher:

     ros2 run p_perf sensor_publisher.py

   • Start the inference node (single or multi-model):

     ros2 run p_perf inferencer.py
     # or
     ros2 run p_perf inferencer_ms.py

   • (Optional) Replay a bag file:

     ros2 run p_perf sensor_replayer.py

4. View results

   • Profiling outputs (timing, GPU/CPU stats) are saved in your specified data directory.
   • Analyze logs and outputs for performance, accuracy, and resource usage.

Demos

For detailed documentation on the experiment scripts, see experiment_scripts/README.md.

To run a demo, simply execute the desired script, e.g.:

python experiment_scripts/bag_test.py

Post-Processing & Analysis

The tools/ directory provides a suite of tools for in-depth analysis of your experiments, including:

• Layer-wise and kernel-level analysis (e.g., analyze_e2e_kernels.py, analyze_memcpy_kernels.py)
• Performance and sensitivity analysis (e.g., performance_analysis.py, dnn_sensitivity_analysis.py)
• Visualization and reporting (e.g., visualize_lidar_scene.py, priority_bar_plots.py)
• Scene and model variation analysis

These tools help you interpret the results, identify bottlenecks, and optimize your models and pipelines.

TODO

• Fix the multi-modal profiling pipeline

Acknowledgments

pPerf builds upon and integrates atop: LISA, nuscenes_to_rosbag, NuScenes, DINO.

Happy profiling! 🚗📊