If you want to access data from one machine on another, then the data has to get to the execute machine somehow.
If you do a lot of remote file access, then the data will still pass across the wires, only in lumps at a time, when you
need it, rather than all up front.
For example, if you have a distributed build system and you want to compile a file called X.c which is 10Mb,
you can copy the whole file across and then compile it, or you could use a networked file system so that the
compiler accesses X.c using networked file I/O without having to copy the whole file. All lines of the file
will still have to be copied across however, so I am sure if you monitored it that 10Mb of data would have flowed across
the wire by the end of the execution.
At least, I am fairly certain thats how it would work for a vanilla universe job.
I/O gets more interesting for standard universe jobs, but there are plenty of experts who can clarify that.