With the increase of the number of nodes in clusters, the probability of failures increases. In this paper, we study the failures in the network stack for high performance networks. We present the design of several fault-tolerance mechanisms for communication libraries to detect failures and to ensure message integrity. We have implemented these mechanisms in the N EW M ADELEINE communication library with a quick detection of failures in a portable way, and with fallback to available links when an error occurs. Our mechanisms ensure the integrity of messages without lowering too much the networking performance. Our evaluation show that ensuring fault-tolerance does not impact significantly the performance of most applications.
