Networks of workstations (NOWs) provide a cost-effective alternative to parallel computers. Components in NOWs may fail, degrading the network operation until the faults are repaired. In this paper, we analyze the influence of both switch and link failures on the network performance. In particular, given that network performance in NOWs strongly depends on the applied routing algorithm, we quantify the sensitivity to failures of two routing algorithms: flexible routing and up*/down* routing algorithms. In the case of up*/down* routing, two methodologies to compute routing tables are evaluated. Evaluation results modeling a Myrinet network show that, in general, up*/down* routing is more robust to failures, although its behavior strongly depends on the type of network topology, regular or irregular, and the methodology used to compute routing tables. However, the flexible routing algorithm presents a better performance, regardless of the network topology, even in presence of failures, but at expense of a larger sensitivity.