In a parallel computing context, peak performance is hard to reach with irregular applications such as sparse linear algebra operations. It requires dynamic adjustments to automatically balance the workload between several processors. The problem becomes even more complicated when an architecture contains processing units with radically different computing capabilities. We present a hierarchical scheduling scheme designed to harness several CPUs and a GPU. It is built on a two-level work stealing mechanism tightly coupled to a software-managed cache. We show that our approach is well suited to dynamically control heterogeneous architectures, while taking advantage of a reduction of data transfers.