I could be wrong, but it was my understanding that, on POSIX systems at least, the 1.8 MRI interpreter was smart enough to use non-blocking IO underneath the hood when performing IO calls and have a thread release its hold on the GIL when it is performing IO. That way, another thread could run while the first thread waits for an IO operation to complete. Either way, whether or not a software-level thread scheduler optimizes by using non-blocking IO or not is orthogonal to whether or not it schedules preemptively or cooperatively.
http://www.igvita.com/2009/05/13/fibers-cooperative-scheduli...
MRI's fair scheduler means a thread gets an equal time slice even if it's blocked on IO.