You would probably need the kernel / filesystem to keep track of it to keep it efficient. And even then that is iffy if you implement it as a kernel extension vs. something integrated into filesystem code.

I've found arq backup overall slow once you start hitting 0.5TB overall. It's a design issue.