This paper presents an efficient MPI implementation on a cluster of PCs using a remote DMA communication primitive. For experimental purposes, the MultiPC (MPC) parallel computer was used. It consists of standard PCs interconnected through a gigabit High Speed Link (HSL) network. This paper focuses on communication software layers over the HSL network. Two implementations of MPI are described. The first one uses hardware interrupts for network events signaling and system calls in the communication critical path. The second one is based on full userlevel communications. Measures show a latency of 15 µs on a Pentium II-350 with this optimized implementation. A quantitative analysis shows how system calls and interrupts impact on communication time. To tally performance in a realistic environment, experiments were run on the Gauss elimination method using a parallel implementation of a local numerical analysis computational package (CADNA).