Application Experiment with the Standard Linux Services for Asymmetric Multiprocessing on Heterogeneous System on a Chips

Modern System on a Chips (SoC) integrate a diverse set of execution units to process the ever-increasing workloads of new embedded applications, i.e., they are heterogeneous SoCs implementing asymmetric multiprocessing. They include general-purpose application processor cores that run an embedded operating system (OS), such as Linux, and they add microcontrollers to execute real-time tasks such as input-output operations, Digital Signal Processors (DSP) for handling stream data such as video or sound, and Graphics Processing Units capable of General-Purpose Compute (GPGPU). These additional execution units, sometimes also called accelerators, are integrated into the software environment utilizing some specialized OS services and applications, generally with some software framework sitting between the OS and the execution units, hiding all the hardware details about the integration of these execution units on the hardware level. GPGPUs are typically handled by specialized frameworks such as CUDA (Compute Unified Device Architecture) or OpenCL (Open Computing Language) fully defining how the execution units are programmed and managed. However, the other types of additional execution units have different, much more diverse HW coupling to the application processors running the OS (such as shared memory, pipes, etc.), and their programming and execution management models can be also fundamentally different. In this paper, we present our experiments with the standard RemoteProc and RPMsg software frameworks of Linux for asymmetric multiprocessing. We did our evaluation on the TI Sitara AM335x and AM57x SoCs platforms utilizing the Beaglebone Black and AI single board computer boards and we also present some initial performance results on the AM335x platform demonstrating the applicability and limits of these technologies.