X's network transparency was basically free at the time
I recently wrote an entry about how X's network transparency has wound up mostly being a failure for various reasons. However, there is an important flipside to the story of X's network transparency, and that is that X's network transparency was almost free at the time and in the context it was created. Unlike the situation today, in the beginning X did not have to give up lots of performance or other things in order to get network transparency.
X originated in the mid 1980s and it was explicitly created to be portable across various Unixes, especially BSD-derived ones (because those were what universities were mostly using at that time). In the mid to late 1980s, Unix had very few IPC methods, especially portable ones. In particular, BSD systems did not have shared memory (it was called 'System V IPC' for the obvious reasons). BSD had TCP and Unix sockets, some System V machines had TCP (and you could likely assume that more would get it), and in general your safest bet was to assume some sort of abstract stream protocol and then allow for switchable concrete backends. Unsurprisingly, this is exactly what X did; the core protocol is defined as a bidirectional stream of bytes over an abstracted channel.
(And the concrete implementation of
$DISPLAY has always let
you specify the transport mechanism, as well as allowing your
local system to pick the best mechanism it has.)
Once you've decided that your protocol has to run over abstracted streams, it's not that much more work to make it network transparent (TCP provides streams, after all). X could have refused to make the byte order of the stream clear or required the server and the client to have access to some shared files (eg for fonts), but I don't think either would have been a particularly big win. I'm sure that it took some extra effort and care to make X work across TCP from a different machine, but I don't think it took very much.
(At the same time, my explanation here is probably a bit ahistorical. X's initial development seems relatively strongly tied to sometimes having clients on different machines than the display, which is not unreasonable for the era. But it doesn't hurt to get a feature that you want anyway for a low cost.)
I believe it's important here that X was intended to be portable across different Unixes. If you don't care about portability and can get changes made to your Unix, you can do better (for example, you can add some sort of shared memory or process to process virtual memory transfer). I'm not sure how the 1980s versions of SunView worked, but I believe they were very SunOS dependent. Wikipedia says SunView was partly implemented in the kernel, which is certainly one way to both share memory and speed things up.
PS: Sharing memory through
mmap() and friends was years in the
future at this point and required significant changes when it arrived.