When running the Deepstream-5.1 on Jetson TX2 and had set the device to 'MAXN' model, but there is always two CPU core not running

Board:Jetson TX2(RAM-8GB/Mem-32GB)
Deepstream-5.1
the other part:opencv/tesorflow/Cuda
we try running the deepstream -5.1 on the Jetson TX2,to analysis the Video which from rtsp link, but we found there is a time delay from the deepstream , The person had left one minute before, but the picture was generated one minute later,it’s due to the Jetson Tx2’s performance not well?there is always two CPU core not running

This is intentional, and you would need to manually schedule tasks to the two Denver cores. You could disable that and use these as general cores.

If you look at “cat /proc/cmdline”, then you will see an argument “isolcpus=1-2” is being passed to the kernel. This tells the scheduler to not schedule tasks there (if you assign something there with taskset, then it will go there even if the scheduler does not do this automatically). To disable this core isolation, see:
https://forums.developer.nvidia.com/t/two-cores-disabled/48637/21

To find more about taskset, see:
https://forums.developer.nvidia.com/t/cannot-enable-denver-cores-for-tx2-jetpack-4-4-dp/124708/40

Hello,Thanks! for this way will improve the performance of the deepstream to analysis the Video?or say how we can improve the performance of the TX2?

I couldn’t predict if this helps or not. For more cores to matter the application must be threaded, and even then, it depends on how you would get cache hits/misses. If two threads are using the same data, then often the scheduler will keep them on the same core since cache hits would go up and cache misses would go down. Perhaps the scheduler would use those other cores, but cache issues would fail to improve anything. In terms of GPU use it wouldn’t change anything for which core(s) the program runs on.

Often, if you were the set up your application to run only on the Denver cores (the core #1-2 of isolcpus), and your program was the only thing using the core, you would possibly see better performance simply because of other apps not causing cache misses. It is really an experiment and it depends on both the code and the data. No way to know without lots of experimentation (and if you want to do things in a truly scientific way, then you’d need to use the profiler tools with each combination of CPU/GPU you test).

Do note that cores #1-2 tend to have more latency than the other cores, but this is not necessarily a problem (it just depends on the nature of the program).

OK, thanks Very much ,Thank you for your feedback

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.