Abstract
We propose a hardware operating system kernel (HOSK), which schedules tasks, controls semaphores, and pre-fetches contexts, as a hardware coprocessor in multiprocessor systems. A multiprocessor system can substantially reduce multithreading overheads by using the HOSK together with simplified RISC processors that do not include hardware for multithreading. We implemented an efficient HOSK which requires about 14 ∼ 25K gates. The experimental results show that the multithreading overheads with a HOSK can be reduced to less than 1 percent. Preliminary efforts confirm that this approach is a feasible solution for minimizing the hardware complexity of a multi-processor system.