On the programmability of multi-GPU computing systems