Unusual performance when working with large messages

Hi, I am getting some strange results when working with large pointcloud messages on a ros2 topic.

With a minimal publisher and subscriber (following this ros tutorial), sending a ~5MB message at 10Hz the listener is picking up maybe %10-20 of messages, however if I startup another subscriber on the same topic, both subscribers will pick up pretty much everything. In this case I was running this code on a VM with 6 cores and 8 GB of RAM, however with more resources, the same effect occurs when adding another subscriber, the only difference is that fewer messages are dropped by the single subscriber. At even higher frequencies I've experienced a single subscriber picking up 0 messages, however two subscribers will still pick up pretty much every message.

How is it that adding subscribers onto a high bandwidth topic improves performance?
I've also run this example using fastdds where most messages are recieved with a single subscriber and adding another subscriber doesn't appear to affect performance.

Start with the linked ROS 2 beginner C++ publisher/subscriber tutorial and reproduce the reported behavior using approximately 5 MB pointcloud messages at 10 Hz. Compare one and two subscribers under Cyclone DDS, then compare with Fast DDS; done means explaining the delivery difference or identifying a focused corrective change.