Implementation and evaluation of a time-driven scheduling processor

Abstract

C.D. Locke developed a heuristic, best-effort (BE) time-driven scheduling policy (Ph.D. thesis, Dept. of Comput. Sci., Carnegie-Mellon Univ., (1986)) and demonstrated its superiority to simpler policies using extensive simulations. The author describes an implementation of Locke's BE policy that uses a dedicated scheduling processor to reduce the scheduling overhead significantly. The scheduler was implemented in the Mach operating system kernel, running on a VAX 11/784 multiprocessor. Experimental results under synthetic real-time processing loads of varying intensity show that the high computational overheads of the BE policy make it impractical on a uniprocessor system. Under heavy loads, over 80% of the processors's time can be spent in the scheduler, deciding which task to run next. Using the scheduling processor, less than 2% of the host processor's time is spent in the scheduler.<>