Using CAPI mode and analysing triton server logs (curl <your_ip>:8002/metrics) :
nv_inference_pending_request_count{model="people_nvidia_detector",version="1"} 0
To sum up, using nveglessink in CAPI mode with several streams, I have latency.
When analysing latency with DeepStream SDK FAQ - #12 by bcao, no significant latency appears. No growing queue at the triton server.
How is it possibly… Can’t find the bottleneck of that latency…