With the emergence of computing infrastructures in the cloud or at the network edge we need to address the question of how to utilize these shared resources when computational tasks are generated dynamically. While small computing tasks may be satisfied with the computing capacity of a single resource, large tasks may want to utilize multiple computing nodes and perform parallel processing to shorten the task completion time. In this paper we evaluate how additional overhead in such divisible load systems affect the efficiency of parallel processing — from the point of view of the task itself, and for the entire resource sharing system. We show that the preference of a single task may be in conflict with the allocation needed for a social optimum, which in turn depends heavily on the load as well as on the system size.