What is more efficient: writing to disk or message passing a huge amount of data as an array between a Python node and CPP node?

edit

My message size is an array (i.e. float32[] of about 648*480*300 or 93312000 elements or 355 MB size). I noticed that writing/reading the array to/from disk takes less time than passing it via a ROS Service. Anyways to speed up the message passing? It takes 5 seconds now to receive the msg.

I would like to note that this slow performance happens even when the service client is set to be persistent.

I'd expect python would take a while to serialize python datatypes, but I wouldn't think it would take 5 seconds to serialize that much data using the optimized numpy methods.