Multi-Protocol Label Switching (MPLS) provides ways to control the Label Switched Paths (LSPs) followed by traffic trunks in a network and thereby to better traffic engineer it. In this context, we look at the problem of organizing the mapping of LSPs in an optimal way throughout the network on the basis of a given objective function. This problem is highly combinatorial and makes dynamic and real-time features a difficult issue for any LSP routing scheme. For this reason, we propose a computationally efficient, though approximate, on-line scheme adapted to an incremental optimization of the network state. It is then applied to a seldom mentioned traffic engineering problem: the compromise between load-balancing and traffic minimization. It is expected that clever routing strategies to balance the network load will sometimes favor longer paths in order to avoid congestion, leading to an increase of the overall network utilization. This reasoning is confirmed by our study, and we show that an improvement in network management can be made by appropriately tuning this compromise.