Building a global time on parallel machines

This paper presents a pragmatic algorithm to buil a global time on any distributed system, which is optimal for homogeneous parallel machines. After some reflexions on time, clocks and distributed systems, we survey and criticize the classical approaches based on clock synchronisation techniques. Satisfying better our purposes, a statistical method is chosen as a building block to derive an original algorithm valid for any topology. This algorithm purposes because, after an acquisition phasis, it induces neither CPU nor message overhead. We provide in the conclusion some data about its behavior and performances on some parallel machines