We investigate the problem of scheduling tasks in a cloud computing scenario that involves a queue of workflow instances whose arrival into the system is governed by a stochastic process. We assume that each instance conforms to the same, known structure, expressed by a directed acyclic graph. Within this job model, the precise execution time of each atomic task and delay of each communication edge may be different and is known only in terms of probability distributions. We argue that in many serial information processing scenarios, minimizing the total response time of the system, i.e., the sum of the execution time and in-queue waiting time, is a more important objective than reducing the schedule length alone, as predominantly done in other approaches. We propose novel algorithms for the minimization of the expected response time and investigate the issue of estimation of the stability of the system. We show that for the given objective, greedy algorithms usually perform better at lower instance arrival rates or when the total amount of instances is limited. However, we also show that an algorithm that minimizes only the expected service time of an instance makes the system more stable and that it must be a non-greedy algorithm.