One of the prime motivators for choosing threads over forks used to be 'overhead'. In older versions of Solaris, for example, the context-switch time between threads was about 1/10 of that for processes. It is still the case that process creation is *much* slower than thread creation. The other was easier mechanisms for interaction (ie, not having to rely on IPC for information-sharing), and overall management (SIGCLD handling is less than elegant).
Now, context-switch is not as big a win as it used to be, but IPC is still more complicated to build and maintain, and is not as portable as pthreads
Of course, this says nothing about the 'challenges' of perl's implementation of threads.