Scaling global scheduling with message passing

Global real-time schedulers have earned the reputation of scaling poorly due to the high runtime overheads involved in global state management. In this paper, two mature implementations, one using fine-grained locking (SCHED DEADLINE) and one using coarse-grained locking (LITMUS
^RT
's G-EDF plugin), are evaluated and it is shown that, regardless of locking granularity, indeed neither scales well w.r.t. worst-case overheads due to excessive lock contention. To demonstrate that this is not an inherent limitation of global scheduling, the design of G-EDF-MP is presented, a global scheduler that uses message passing to avoid lock contention and cache-line sharing. It is shown to offer up to a 23- to 36-fold reduction in worst-case scheduling overhead on a 64-core platform, which translates into much improved schedulability (in some cases, more than 120 additional tasks can be supported).