Operating System Process and Thread Migration in Heterogeneous Platforms

Parallel and heterogeneous computing are here to stay. Moreover, diverse computational units are increasingly tighter-integrated in emerging heterogeneous platforms, sharing access to the memory bus. We argue that the traditional way we use heterogeneous platforms is obsolete. We propose newsystem software that enables programming these platforms as if they were an SMP, via the multi-threaded shared-memory programming model. Using this model provides better programmability, flexibility, and exploitability. We first provide an overview of the state transformation problem when migrating a thread between different-ISA processors. We then propose, build and evaluate a series of offline and runtime mechanisms that implement our design on top of Popcorn Linux ARM/x86. Results show that the overheads are minimal, thus proving the viability of our architecture.