The purpose of this paper is to investigate how to design an integrated routing architecture for IP and ATM meeting the requirements for a large scale Internet based on IP and ATM. Integration of IP and ATM at the routing level leads us to consider two separate aspects: using a common routing architecture for IP and ATM on one hand (layer integration) and, on the other hand, integrating best-effort and QoS traffic support in the same routing architecture (service integration). The first level of integration is, for obvious reasons, highly recommended. In contrast, we show that the second level of integration is not desirable because best- effort and QoS traffic flows have, in terms of routing contradictory requirements. To conduct this analysis, we feel that, because of the inherent complexity of the problem, confronting the existing proposals is too restrictive. Instead, we propose to go one step back in the design process and identify the basic design options to be considered when designing a routing architecture. We identify three options, namely, route updating vs. route pinning, hop by hop vs. explicit routing and pre-computed routes vs. on-demand route computation. A fourth option is whether or not to integrate in the routing architecture the capability to compute shortcut paths, that is, bypassing layer 3 (L3) nodes and using only layer 2 (L2) devices. Using this framework, we conclude that best-effort traffic flows are well served by a combination of route updating, hop by hop routing and pre-computed routes while QoS flow routing is built on route pinning, explicit routing and on-demand route computation. We also observe that the capability to compute L2 shortcuts in an L2/L3 integrated routing architecture is an added value simplifying the overall network design and optimizing the efficiency of the forwarding path.