The queue is empty when processing videos

Accelerated Computing Intelligent Video Analytics DeepStream SDK
,
1

• Hardware Platform (Jetson / GPU): GPU
• DeepStream Version: 7.1 (python)
• TensorRT Version: 10.3.0.26
• NVIDIA GPU Driver Version (valid for GPU only): 555.58.02
• Issue Type( questions, new requirements, bugs): questions
Hello. I have a few questions about working with the queue.

  1. Where are the frames that are in the queue stored? Are they on RAM or GPU memory?
  2. How do I see how many frames are in the queue? In python deepsteam. I used .add_probe for src and sink pads, and looked at the attributes current-level-time, current-level-bytes. But they always returns 0. For reading, I use nvmultiurisrcbin, and after that there is a queue and then nvinferserver.
    изображение
    изображение692×499 45.2 KB

    Screenshot from 2024-12-23 16-01-08
    Screenshot from 2024-12-23 16-01-08692×188 19 KB
3

The answer to this question depends on the type of your GstBuffer. If it is an NVMM type buffer, the video frame is stored on the GPU memory, but the nvbufsurface is stored as a handle on the CPU. Other types of buffers are stored in RAM.

It is impossible to know exactly, current-level-bytes is 0, which means there is no delay in data processing.

4

Thanks for the reply. As for the problem with the current-level-bytes is 0. There must be a delay in processing the data, as messages on the triton server take a very long time to be sent. Accordingly, because of this, the queue should fill up very quickly, since nvmultiurisrcbin continues to read frames.

5

Please tell me how I can solve the problem with the fact that nothing adds up to the queue.

6

I don’t understand your intention. Besides, this question has nothing to do with DeepStream.

In fact, I tried your pipeline and got the following results

current_level_bytes 64
current_level_bytes 0
current_level_bytes 0
current_level_bytes 0
current_level_bytes 0
NvMMLiteOpen : Block : BlockType = 261 
NvMMLiteBlockCreate : Block : BlockType = 261 
current_level_bytes 0
current_level_bytes 0
current_level_bytes 64

deepstream_test1_app.c (11.1 KB)

A queue is the thread boundary element through which you can force the use of threads. Although triton has a delay, it does not mean that data will be cached in the queue

7

Thanks for the reply. And how, then, is it possible to emulate the accumulation of data in a queue?

8

You can simulate a slow sink， On my device, I can get the following logs:

Queue buffer count: 6
Queue buffer count: 6
Queue buffer count: 6
Queue buffer count: 6
Queue buffer count: 6
Queue buffer count: 6

import gi
gi.require_version('Gst', '1.0')
from gi.repository import Gst, GLib

Gst.init(None)

def on_pad_probe(pad, info, user_data):
    queue = user_data
    current_level_buffers = queue.get_property("current-level-buffers")
    print(f"Queue buffer count: {current_level_buffers}")
    return Gst.PadProbeReturn.OK

pipeline = Gst.parse_launch("videotestsrc ! video/x-raw,width=1920,height=1080,framerate=100/1 ! queue max-size-buffers=10 min-threshold-buffers=0 name=q0 ! ximagesink sync=true processing-deadline=1000000")

queue = pipeline.get_by_name("q0")
if not queue:
    print("Could not find queue element")
    exit(-1)

srcpad = queue.get_static_pad("src")
srcpad.add_probe(Gst.PadProbeType.BUFFER, on_pad_probe, queue)

pipeline.set_state(Gst.State.PLAYING)

try:
    loop = GLib.MainLoop()
    loop.run()
except KeyboardInterrupt:
    pass
finally:
    pipeline.set_state(Gst.State.NULL)
9

Can you share your goals? I might be able to give you better advice.

13

Apologize for the long answer. I have a deep stream pipeline that processes live video streams. And I would like to watch on these cases when there is a delay between the reader module and inference module. Between these two modules I have a queue (nvmultiurisrcbin-> queue->nvinferserver) .

Which, if I understand correctly, just accumulates frames in the event that nvinferserver does not have time to process later frames?
And I wanted to keep track of how many frames are currently in the queue. Can you tell me if this is possible?

14

Can you help me please?

15

1.Refer to this FAQ for get element latency.

You can try using nvinferserver grpc mode, and then try to get tritonserver Pending Request Count

How to use nvinferserver in grpc mode, please refer to the following README

/opt/nvidia/deepstream/deepstream/samples/configs/deepstream-app-triton-grpc/README

https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/user_guide/metrics.html

 curl localhost:8002/metrics | grep nv_inference_pending_request_count
16

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.