If you are adventurous enough, you could try threads, but using them is "officially discouraged" (though not (yet) deprecated).

Lowering the read size, while very simple, will significantly degrade performance.

Non-blocking writes will add a lot of complexity. I think you will have to select on writability as well as readability.

If there is known maximum number of clients, you could pre-fork "worker" processes. Your "master" process would listen for connect requests, then "hand" them off to the next available worker. Unfortuantely, it has been many years since I did something like that, so I don't remember how to do the "hand off".

Unless there are performance or resource reasons to not use the fork-on-demand model, is probably best to just stay with that.