This paper presents a study of the interaction between high-speed interconnects and a distributed file system client. We use our remote file access protocol and network with the software layer to show how the highly specific programming model of high-speed interconnects may be used in a high performance distributed file system client. We present both a user-level and a kernel-level client implementations with either buffered or non-buffered accesses. These implementations show high performances and may be transparently used by applications. Our improvements focus on pin-down cache techniques and memory registration issues. Our modifications show no impact on its performances while the network usage in the distributed file system client is improved.