"... manipulate an existing csv file without reading it (and without saving the new results). Is this possible in matlab?"
Taken literally, this is impossible without the yet-to-be-released Clairvoyant Toolbox™.
Practically, what you're describing is to simply keep the data in a shared memory area. I've never had the ML parallel programming TB so don't know how it deals with it; with what I have done one uses semaphores to handle the access to the shared data from the different processes -- gridlock is a potential issue when the number of processes gets large, I would suspect, but I've never worked in anger with more than a couple processors at a time so new techniques are bound to have been developed since...
