In this paper a semi-online algorithm for scheduling multiprocessor tasks with partial information is proposed. We consider the case in which it is possible to exploit probabilistic information and use this information to obtain better solutions in comparison with standard non clairvoyant on-line algorithms. A wide computational analysis shows the effectiveness of our algorithm. Moreover, we also consider a test framework with a continuous generation of tasks in order to study the behavior of the proposed approach in real applications, which confirms the efficiency of our approach.