One interpreter per-thread means all side-effects have to be migrated to the other threads to keep a consistent view of memory: guess what you will need to do that? yep, a global lock (except this time it's across all interpreters, instead of just one.)
For scientific computing purposes, you can often accomplish this with multiple processes and a numpy array allocated by shmget/shmat. But I'm not sure how to share complex objects in this way.
But I think that's how the Queue object in Python 2.6 works. The Queue instance is locked, you seem to be free to do whatever within the threads that are consuming the queue.
The reason I'm not sure is that having a single object being locked seems to contradict the GIL concept...
I like perl's ithread setup a lot. You explicitly need to mark data as shared between threads, otherwise all variables/objects are local to the thread. Things like queues, for example, are implemented as shared @arrays with a locking primitive around the manipulations, mostly hidden behind the Queue API (Queue->enqueue and Queue->dequeue).
The interpreter code is still shared among all the threads, but each has thread local storage separate from the shared arena. I've found that explicitly needing to mark data as shared to be a big boon to development, I think it has helped reduce the number of the bugs related to shared state.
I can't produce a significantly long start up time when perl copies a 240meg+ local state using http://gist.github.com/258016. Much larger, and the overhead pushes my 4gig machine into swap (so perl's not very efficient with its data overhead, it seems). It is slower than if the structure is shared (you can uncomment a line to compare) between the threads, which shows that perl is copying the state to the other interpreters, but this just says you need to use the same model with perl threads that you'd use when you do multiprocessing with fork if you're going to have long lived child processes that don't exec: spawn threads early before your heap balloons to avoid all that spawn-time copying (which is a good practice anyway, even when COW is used for fork). Every model has its trade offs, and it's good to know those trade offs going in, rather than treating the flavor of the day as a silver bullet.
And even if perl's implementation isn't stable/good, it doesn't mean the model of explicit sharing is a bad one. And it most definitely doesn't mean that using a separate interpreter in each thread with global sharing is bad at all.
vfork() and clone() can almost give you what you want, but Linux threading is using something different (NPTL) nowadays.
Even with NPTL, both processes and threads have task_structs.