subscribe image slow on rospy with python3 and melodic

edit

I've seen several issues posted about rospy being slow to subscribe to images in particular:

https://answers.ros.org/question/5011... https://answers.ros.org/question/2967... https://answers.ros.org/question/2123... https://answers.ros.org/question/2205... https://answers.ros.org/question/3629... https://github.com/ros2/ros2/issues/509

On Jetson Xavier NX with L4T (ARM) image 32.4.4, I've tried all of the advice listed in those issues, namely:

1. setting PYTHONOPTIMIZE=0 env var

2. using queue_size for asynchronous subscriber

3. Using buffer size to buffer large images completely (1280x720 raw images at ~24 fps):

   rospy.Subscriber(image_topic, Image, image_callback, queue_size=1, buff_size=100428800)

4. Connecting over ethernet to the other machine publishing the images at 24 fps (rosbag)

Still I am only achieving a max of 2.5 FPS subscription rate, which is unacceptable for real-time image processing. Any other suggestions for how to speed up subscribing images?

my image callback is just logging the FPS. In roscpp subscribing to this same image topic does not have a significant lag. Am I hitting a fundamental limit of python(3)?

# all code besides ros spin and imports. Topic published by rosbag  as 1280x720 rgb raw image
def image_callback(self, data):
        current = rospy.get_time()
        rospy.loginfo(f"image FPS: {1/(current - self.last_time)}")
        self.last_time = current

Note that I had to build ros from source setting python executable to python3 in order to use python3 packages with ros melodic.