Stream Performance Drops When Number of Streams Less than Batch Size

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Please provide complete information as applicable to your setup.

**• GPU (RTX 3060 Ti)
**• DeepStream Version: 6.2
**• TensorRT Version: 8.6.1.6
• NVIDIA GPU Driver Version (valid for GPU only) 525.105.17
• Issue Type( questions, new requirements, bugs) Question
• How to reproduce the issue ? (This is for bugs. Including which sample app is using, the configuration files content, the command line used and other details for reproducing)

  1. Start the deepstream-app reference application, configured with a streammux batch-size of, say, 4, and further configure the application to use anywhere between 1-3 (inclusive) RTSP streams, preferably streams with known frame rates (I’m running 30fps 1080p streams).
  2. Watch the performance metrics. They all show ~25fps (at least in my case).
  3. Restart the deepstream-app reference application but this time with a streammux batch-size matching the number of RTSP streams.
  4. Watch the performance metrics. They all show 30fps.

This is a bit confusing to me as the docs for nvds-streammux say “batch-size” should be the maximum size of the batches to be processed, and so one would presume that a pipeline with fewer sources than “batch-size” would perform just fine.

In fact, I’ve noticed that no matter the batch size, if I use any number of streams smaller than the batch size, I always see a 5fps drop in each stream.

I also put some debug code into deepstream_perf.c. Nothing looks out of the ordinary, but the corresponding “buffer_cnt” property in NvDsInstancePerfStruct has a value of 25 less in the case where the number of streams is smaller than the batch size.

E.g. buffer_cnt == 125 when the performance timer callback fires every 5s when the number of streams is fewer than the batch size, and buffer_cnt == 150 otherwise.

Possibly related, based on https://developer.nvidia.com/blog/managing-video-streams-in-runtime-with-the-deepstream-sdk/, I would think that updating batch-size in the appropriate plugins would be necessary; however, I’ve tried doing something similar by making adjustments to the deepstream-app application, but I’ve noticed that if I start the pipeline with, say, 4 streams (with batch-size set to 4), and then later dynamically remove 1 or more streams while also updating the streammux with the appropriate batch-size value, I see the same 5fps drop.

Any help or insight would be greatly appreciated.

Thank you.

3
  1. Have you set “live-source” property of nvstreammux as 1 for live sources? Gst-nvstreammux — DeepStream documentation 6.4 documentation
  2. Have you followed the instruction DeepStream SDK FAQ - Intelligent Video Analytics / DeepStream SDK - NVIDIA Developer Forums to set “batched-push-timeout” property of nvstreammux?
4

Thanks, #2 was the answer (I had already done #1).

Given how common this issue looks to be (I can see similar issues on this forum but phrased differently so as not to make it immediately obvious), I’m a bit surprised it isn’t linked from the streammux documentation.

Regardless, thank you for your your help!