Résumé : We propose a multicore-ready programming language based on a two-level shared memory model. Concurrency units are schedulers and threads which are dispatched on available cores in a preemptive way. Each scheduler is in charge of its own portion of the memory. At runtime, several threads may link to a common scheduler. In this case, they enter a cooperative mode, evolve in synchronous rounds, and are granted access to the scheduler memory. At the opposite, an autonomous thread runs at its own pace but has access only to a local memory. The language ensures that programs are free of memory leaks, that code between two cooperation points is atomic, and that rounds are fair and always terminate (no run-time error nor divergence).