I have a question about viewing mapped memory on CUDA (GPU side).
For CPU side, We can see /proc/PID/maps etc on Linux. And break down into memory usage with various tools.
Is there any tool to see memory usage on CUDA (GPU side)? At least, viewing each CUDA kernel code meory usage is useful.
Using nvidia-smi you can get an instantaneous snapshot of GPU process memory usage, i.e. how much GPU memory is being used by that process.
Using a profiler (e.g. nsight compute) you can get measurements of how much read and write activity to global memory is occurring from a kernel.
Thank you for your comments.
As you suggested, nvidia-smi can monitor total memory usage.
I want to know which function uses huge memory size. Is there any method to monitor this?