Shared libraries per se should not be a problem (TBB comes as a shared library, and even treats the scalable allocator as a dynamically loadable library), but obviously you shouldn't look up a function by name from a dynamically loadable library on each invocation (shared_library::get() from Boost.Extension, or similar). If I can assume that "each function is a class loaded from a shared library" means that your tbb::function_node Body calls a function object from a dynamically loadable library, isn't the question now about the performance of boost::function after it has been assigned a value? There's obviously some run-time indirection going on there, strictly speaking, but you needn't worry about that here (as opposed maybe to calling a minimal function in a tight loop). Maybe it's even possible to load a subclass of an abstract Body (right?), and then you would know it's a straight virtual-function call (not necessarily cheaper than a boost::function call, but at least more familiar), but I would have to check that first.
