Managing tardiness bounds and overload in soft real-time systems.

摘要

In some systems, such as future generations of unmanned aerial vehicles (UAVs), different software running on the same machine will require different timing guarantees. For example, flight control software has hard real-time (HRT) requirements---if a job (i.e., invocation of a program) completes late, then safety may be compromised, so jobs must be guaranteed to complete within short deadlines. However, mission control software is likely to have soft real-time (SRT) requirements---if a job completes slightly late, the result is not likely to be catastrophic, but lateness should never be unbounded.;The global earliest-deadline-first (G-EDF) scheduler has been demonstrated to be useful for the multiprocessor scheduling of software with SRT requirements, and the multicore mixed-criticality (MC2) framework using G-EDF for SRT scheduling has been proposed to safely mix HRT and SRT work on multicore UAV platforms. This dissertation addresses limitations of this prior work.;G-EDF is attractive for SRT systems because it allows the system to be fully utilized with reasonable overheads. Furthermore, previous analysis of G-EDF can provide "lateness bounds" on the amount of time between a job's deadline and its completion. However, smaller lateness bounds are preferable, and some programs may be more sensitive to lateness than others. In this dissertation, we explore the broader category of G-EDF-like (GEL) schedulers that have identical overhead characteristics to G-EDF. We show that by choosing GEL schedulers other than G-EDF, better lateness can be achieved, and that certain modifications can further improve lateness bounds while maintaining reasonable overheads. Specifically, successive jobs from the same program can be permitted to run in parallel with each other, or jobs can be split into smaller pieces by the operating system.;Previous analysis of MC2 has always used less pessimistic execution time assumptions when analyzing SRT work than when analyzing HRT work. These assumptions can be violated, creating an overload that causes SRT guarantees to be violated. Furthermore, even in the expected case that such violations are transient, the system is not guaranteed to return to its normal operation. In this dissertation, we also provide a mechanism that can be used to provide such recovery.