DeepStream RTSP FPS Drop in Docker – Performance Issue

[user111099]

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson / GPU) - GPU (A4000)
• DeepStream Version - 7.1
• TensorRT Version - 10.8.0.43-1
• NVIDIA GPU Driver Version (valid for GPU only) - 550

I’m experiencing a significant FPS drop while running multiple RTSP streams using DeepStream in Docker. Below is a brief of my setup:

Docker Image: nvcr.io/nvidia/deepstream:7.1-gc-triton-devel

Streams: 8 RTSP streams

Pipeline:

streammux → pgie → tracker → sgie1 → sgie2 → sgie3 → nvvidconv1 → filter1 → nvanalytics → tiler → nvvidconv → nvosd → tee
tee → queue1 → msgconv → msgbroker
→ queue2 → nvvidconv_postosd → caps → encoder → rtppay → sink

At the start, I get around 21–22 FPS per stream, but over time it drops drastically to 5–6 FPS.

But, when I run the same pipeline locally (outside Docker), I start at 25–27 FPS per stream and it gradually drops to 17–18 FPS.

How can i debug the bottleneck?

[Fiona.Chen]

Have you monitored the GPU loading and CPU loading when the FPS dropped?

There are 3 SGIEs in your pipeline, the SGIEs may run on objects detected by PGIE, have you measured the models’ performance on A4000?

[user111099]

[Fiona.Chen]

You can measure the model’s performance with “trtexec” tool.
You may refer to Low performance when running pipeline with RTX 4090 - Intelligent Video Analytics / DeepStream SDK - NVIDIA Developer Forums