Improved analysis and evaluation of real-time semaphore protocols for P-FP scheduling

Several suspension-based multiprocessor real-time locking protocols for partitioned fixed-priority (P-FP) scheduling have been proposed in prior work. These protocols differ in key design choices that affect implementation complexity, overheads, and worst-case blocking, and it is not obvious which is “best” when implemented in a real OS. In particular, should blocked tasks wait in FIFO or in priority order? Should tasks execute critical sections locally on their assigned processor, or should resource access be centralized on designated processors? This paper reports on a large-scale, overhead-aware schedulability study comparing four protocols, the MPCP, FMLP
⁺
, DPCP, and the DFLP, which together cover each of the four possible combinations. The results are based on a new, linear-programming-based blocking analysis technique, which is explained in detail and shown to offer substantial improvements over prior blocking bounds. The results reveal that priority queuing (MPCP, DPCP) is often preferable if the range of temporal constraints spans (at least) an order of magnitude, whereas FIFO queueing (FMLP
⁺
, DFLP) is preferable if the ratio of longest to shortest deadlines is small. Further, centralized resource access (DPCP, DFLP) is found to be preferable to local critical sections (MPCP, FMLP
⁺
) for high-contention workloads. Scheduling, cache, and locking overheads were accounted for as measured in LITMUS
^RT
on two 8- and 16-core x86 platforms. In contrast to earlier LITMUS
^RT
-based studies, no statistical outlier filtering was performed, owing to improved tracing support.