Optimizing network slice placement in 5G networks requires efficient algorithms. Deep Reinforcement Learning (DRL) has been used to solve this problem successfully. However, few works have tackled the deployment of these algorithms in a real environment. In this paper we present a DRL based algorithm aiming to optimally place network slices in IoT networks. We evaluate the performance of this algorithm in a real network deployed on Grid'5000 platform, operated by an Open Source MANO (OSM) middleware. The simulation results show a good convergence of the algorithm and the deployment in the real environment gives us some insights about a potential slicing architecture using OSM, the processing of a DRL agent in real conditions, and limitations due to consequent instantiation times for Virtual Network Functions (VNF).