I want to find the GPU memory amount by each process on a jetson nano or a similar board. Similar to the command:
nvidia-smi --query-compute-apps=pid,used_memory --format=csv
But Since the nvidia-smi is not available, I cannot do the same.
Is there a tool or a library that can find the memory per process?
Jetsons have an iGPU (as opposed to dGPU) sharing same memory as CPUs. So there is not GPU memory, just some memory that may be used from CPU or GPU or both (the latter requires using Unified or ZeroCopy memory).
So you would just look at memory used by each process as a whole.
Thanks for your reply. However, All programs that read host memory doesn’t include the memory allocated for the GPU. For instance, create a program that uses CUDA and measure the memory by psutil and check the total allocated memory using free before and during the program. You will find a difference in hundreds of GPS (depending on the program). I was already using this method to capture the program’s total memory and know the capacity of each process individually. But, when I switched to the multiprocess scenario doing this wasn’t effective.
Maybe I spoke too fast, but looking at htop results (sudo apt install htop) when using GPU allocated memroy, I do see some difference.
Someone more skilled with this may further advise.