Soft real-time scheduling on multiprocessors

The design of real-time systems is being impacted by two trends. First, tightly-coupled multiprocessor platforms are becoming quite common. This is evidenced by the availability of affordable symmetric shared-memory multiprocessors and the emergence of multicore architectures. Second, there is an increase in the number of real-time systems that require only soft real-time guarantees and have workloads that necessitate a multiprocessor. Examples of such systems include some tracking, signal-processing, and multimedia systems. Due to the above trends, cost-effective multiprocessor-based soft real-time system designs are of growing importance. Most prior research on real-time scheduling on multiprocessors has focused only on hard real-time systems. In a hard real-time system, no deadline may ever be missed. To meet such stringent timing requirements, all known theoretically optimal scheduling algorithms tend to preempt process threads and migrate them across processors frequently, and also impose certain other restrictions. Hence, the overheads of such algorithms can significantly reduce the amount of useful work that is accomplished and limit their practical implementation. On the other hand, non-optimal algorithms that are more practical suffer from the drawback that their validation tests require workload restrictions that can approach roughly 50% of the available processing capacity. Thus, for soft real-time systems, which can tolerate occasional or bounded deadline misses, and hence, allow for a tradeoff between timeliness and improved processor utilization, the existing scheduling algorithms or their validation tests can be overkill. The thesis of this dissertation is: Processor utilization can be improved on multiprocessors while providing non-trivial soft real-time guarantees for different soft real-time applications, whose preemption and migration overheads can span different ranges and whose tolerances to tardiness are different, by designing new algorithms, simplifying optimal algorithms, and developing new validation tests. The above thesis is established by developing validation tests that are sufficient to provide soft real-time guarantees under non-optimal (but more practical) algorithms, designing and analyzing a new restricted-migration scheduling algorithm, determining the guarantees on timeliness that can be provided when some limiting restrictions of known optimal algorithms are relaxed, and quantifying the benefits of the proposed mechanisms through simulations. First, we show that both preemptive and non-preemptive global earliest-deadline-first(EDF) scheduling can guarantee bounded tardiness (that is, lateness) to every recurrent real-time task system while requiring no restriction on the workload (except that it not exceed the available processing capacity). The tardiness bounds that we derive can be used to devise validation tests for soft real-time systems that are EDF-scheduled. Though overheads due to migrations and other factors are lower under EDF (than under known optimal algorithms), task migrations are still unrestricted. This may be unappealing for some applications, but if migrations are forbidden entirely, then bounded tardiness cannot always be guaranteed. Hence, we consider providing an acceptable middle path between unrestricted-migration and no-migration algorithms, and as a second result, present a new algorithm that restricts, but does not eliminate, migrations. We also determine bounds on tardiness that can be guaranteed under this algorithm. Finally, we consider a more efficient but non-optimal variant of an optimal class of algorithms called Pfair scheduling algorithms. We show that under this variant, called earliest- pseudo-deadline-first (EPDF) scheduling, significantly more liberal restrictions on workloads than previously known are sufficient for ensuring a specified tardiness bound. We also show that bounded tardiness can be guaranteed if some limiting restrictions of optimal Pfair algorithms are relaxed. The algorithms considered in this dissertation differ in the tardiness bounds guaranteed and overheads imposed. Simulation studies show that these algorithms can guarantee bounded tardiness for a significant percentage of task sets that are not schedulable in a hard real-time sense. Furthermore, for each algorithm, conditions exist in which it may be the preferred choice