Hi all.
I have been testing a few GStreamer pipelines in order to check the memory usage in create-start-stop-delete sequences. I am using RidgeRun’s Gstreamer Daemon (gstd) to perform such operations on the pipelines.
In this case I focused on the nvvideoconvert + nvv4l2h264enc combination.
Setup details:
• Jetson TX2
• DeepStream 5.0.0
• JetPack 4.4.3
• TensorRT 7.1.3
Here is the very basic script I am using:
#!/bin/bash
function interrupt() {
echo
echo "Ctrl-C interruption"
rm -rf ./data/
gstd-client pipeline_delete p1
exit
}
i="0"
while [ $i -lt 100 ]
do
echo "Creating p1"
gstd-client pipeline_create p1 videotestsrc ! nvvideoconvert ! "video/x-raw(memory:NVMM)" ! nvv4l2h264enc insert-sps-pps=true iframeinterval=5 control-rate=1 profile=4 qos=false maxperf-enable=true name=h264_encoder ! fakesink
echo "Playing p1"
gstd-client pipeline_play p1
sleep 10
echo "Stopping p1"
gstd-client pipeline_stop p1
gstd-client pipeline_delete p1
sleep 8
i=$[$i+1]
done
It runs the create-play commands, waits for 10 seconds, runs the stop-delete commands, waits for 8 seconds and repeats the cycle. Each test was performed during 400 seconds.
Results:
I am also using a script that runs pmap to extract the shared, private and mapped memory of a specific process (gstd in this case). This script also provides plots of the memory usage, which are presented below as part of the results.
- Pipeline A:
nvvideoconvert+nvv4l2h264enc
gstd-client pipeline_create p1 videotestsrc ! nvvideoconvert ! "video/x-raw(memory:NVMM)" ! nvv4l2h264enc insert-sps-pps=true iframeinterval=5 control-rate=1 profile=4 qos=false maxperf-enable=true name=h264_encoder ! fakesink
Private Memory:
Mapped Memory:
Shared Memory:
- Pipeline B:
nvvideoconvertonly
gstd-client pipeline_create p1 videotestsrc ! nvvideoconvert ! "video/x-raw(memory:NVMM)" ! fakesink
Private Memory:
Mapped Memory:
Shared Memory:
I would appreciate any advice or suggestion on how to avoid this behavior, including any property that should be set.
Thank you in advance,
-JC