In this article, we study a method to compare queueing systems and their fl
uid limits. For a certain class of queueing systems, it is shown that the e
xpected workload (and certain functions of the workload) is higher in the q
ueueing system than in the fluid approximation. This class is characterized
by convexity of the value function in the state component(s) where externa
l arrivals occur. The main example that we consider is a tandem of multiser
ver queues with general service times and Markov-modulated arrivals. The an
alysis is based on dynamic programming and the use of phase-type distributi
ons. Numerical examples to illustrate the results are also given.