Because a delay tolerant network (DTN) can often be partitioned, routing is a challenge. However, routing benefits considerably if one can take advantage of knowledge concerning node mobility. This paper addresses this problem with a generic algorithm based on the use of a high-dimensional Euclidean space, that we call MobySpace, constructed upon nodes' mobility patterns. We provide here an analysis and a large scale evaluation of this routing scheme in the context of ambient networking by replaying real mobility traces. The specific MobySpace evaluated is based on the frequency of visits of nodes to each possible location. We show that routing based on MobySpace can achieve good performance compared to that of a number of standard algorithms, especially for nodes that are present in the network a large portion of the time. We determine that the degree of homogeneity of node mobility patterns has a high impact on routing. And finally, we study the ability of nodes to learn their own mobility patterns.