In this paper we compare two traffic models based on Markov modulated Poisson processes (MMPPs), that were designed to capture self-similar behavior over multiple time scales. These models are both constructed by fitting the distribution of packet counts in a number of time scales. The first model is a superposition of MMPPs where each MMPP describes a different time scale. The second one is obtained as the equivalent to an hierarchical construction process that, starting at the coarsest time scale, successively decomposes MMPP states into new MMPPs to incorporate the characteristics offered by finner time scales. We evaluate the accuracy of the models by comparing the probability mass function at each time scale, as well as the loss probability and average waiting time in queue, corresponding to measured traces and to traces synthesized according to the proposed models. The analysis is based on three measured traffic traces exhibiting self-similar behavior: the well-known pOct Bellcore trace and two traces measured in a Portuguese ISP. Based on the obtained results, we conclude that both Markovian models have good and very similar performances in matching the characteristics of the data traces over the relevant time scales. However, one advantage of the hierarchical approach is that the number of states of the corresponding MMPP can be much smaller.