not only when, but where : a fairly common pattern is to have computations done in one thread, then once the computation is done and the thread starts another, the shared_ptrs (or any other kind of memory owners) are swapped in another thread and cleared there.