Many applications in data center networks are very sensitive to delays. To reduce delay variability, one may introduce priorities in the network switches and the servers. The priorities can be based on characteristics of the flows, such as a class or a residual file length. This paper considers another approach where all the regulation is performed end-to-end, with the simplest possible switches. The study indicates that this simpler architecture enables low delay variability and high utilization.  The algorithms are based on a new variation of the network optimization problem.