A computer system is modeled by an exponential queueing network with different classes of customers and a general class of service policies. The mean cost per unit time is taken as the loss function. Three lower bounds of the loss function under the whole class of service policies, as well as an upper bound of the minimal loss, are derived. The loss function is expressed in a convenient form from which we derive a simple heuristic service policy so called "Klimov policy." This policy is applied to several examples of computer systems and is evaluated by the bounds. The examples recommend using "Klimov policy" in any case; so we have a first rule for deciding how to provide service capacity to different customers in a network of queues.