To tackle the multi-core programming challenge, we investigate the design and implementation of concurrency-oriented programming languages. Our approach mimics the evolution from lambda-calculi to functional programming languages, but with the pi calculus as a starting point. To fill the gap between the abstract calculus and its implementations, we introduce the pi-threads: an intermediate language and its abstract machine.The stackless architecture of the abstract machine makes the underlying algorithms both simple and naturally concurrent. The scheduling, for instance, can be operated in a completely decentralized way. Another remarkable feature of the abstract machine is its garbage collector. We adopt a reference counting scheme that can be characterized formally using only two semantic rules. Moreover, it provides original solutions to the usual shortcomings of reference counting: the overhead caused by the maintenance of the reference counts - we only track global references - and the complex issue of collecting cyclic structures - reinterpreted as the (in our case, much simpler) problem of detecting partial terminations.