As the performance demands imposed on relational database systems are steadily increasing, it becomes more difficult to master then with conventional database architectures. New architectural proposals suggesting the use of intelligent disc subsystems in combination with appropriate set-oriented interfaces look promising, because they can be realised by standard, high-performance hardware now. In particular, it makes sense to exploit the fast sequential read-out capabilities of modern discs. In such a changed environment the access-path selection problem, especially the question of whether index usage can speed up retrieval query execution, requires new solution approaches. Based on easily acquired performance data, an analytical model for restriction queries is developed that allows us to determines threshold hit ratios separating profitable from non-profitable index usage. As this model is founded on an extent-based file organisation, it is likewise applicable to conventional architectures permitting chained I/O. Simulation results for two common disc-pack drives are reported, showing that the bottleneck for random disc accesses sharpens for modern discs, and thus makes exhaustive relation scans preferable in more cases. However, indexes will still be mandatory on highly selective attributes to achieve satisfactory performance. The presented optimisation criteria are directly amenable to incorporation into a query optimiser.