Non-constant inference time with pycaffe

I have a python script that reads images from a USB camera and feeds them to a Caffe neural network. I tried to log the inference and acquisition time at each iteration and I noticed that there are some strange peaks that I don’t know how to explain. I tried to set the number of cores using taskset (all cores, two A72, one A72) but the issue remains. I am attaching the plot for the 2 cores case (I can only upload one image).

I think the issue is on the USB side? We have increased the new user max image number. You can post the other images as well.