On Best-Effort Utility Accrual Real-Time Scheduling on Multiprocessors

Abstract

We consider the problem of scheduling real-time tasks on a multiprocessor system. Our primary focus is scheduling on multiprocessor systems where the total task utilization demand, U , is greater than m, the number of processors on a multiprocessor system-i.e., the total available processing capacity of the system. When U > m, the system is said to be overloaded; otherwise, the system is said to be underloaded. While significant literature exists on multiprocessor real-time scheduling during underloads, little is known about scheduling during overloads, in particular, in the presence of task dependencies-e.g., due to synchronization constraints. We consider real-time tasks that are subject to time/utility function (or TUF) time constraints, which allow task urgency to be expressed independently of task importance-e.g., the most urgent task being the least important. The urgency/importance decoupling allowed by TUFs is especially important during overloads, when not all tasks can be optimally completed. We consider the timeliness optimization objective of maximizing the total accrued utility and the number of deadlines satisfied during overloads, while ensuring task mutual exclusion constraints and freedom from deadlocks. This problem is NP-hard. We develop a class of polynomial-time heuristic algorithms, called the Global Utility Accrual (or GUA) class of algorithms.The algorithms construct a directed acyclic graph representation of the task dependency relationship, and build a global multiprocessor schedule of the zero in-degree tasks to heuristically maximize the total accrued utility and ensure mutual exclusion. Potential deadlocks are detected through a cycle-detection algorithm, and resolved by aborting a task in the deadlock cycle. The GUA class of algorithms include two algorithms, namely, the Non-Greedy Global Utility Accrual (or NG-GUA) and Greedy Global Utility Accrual (or G-GUA) algorithms. NG-GUA and G-GUA differ in the way schedules are constructed towards meeting all task deadlines, when possible to do so. We establish several properties of the algorithms including conditions under which all task deadlines are met, satisfaction of mutual exclusion constraints, and deadlock-freedom.We create a Linux-based real-time kernel called ChronOS for multiprocessors. ChronOS is extended from the PREEMPT RT real-time Linux patch, which provides optimized interrupt service latencies and real-time locking primitives. ChronOS provides a scheduling framework for the implementation of a broad range of real-time scheduling algorithms, including utility accrual, non-utility accrual, global, and partitioned scheduling algorithms.We implement the GUA class of algorithms and their competitors in ChronOS and conduct experimental studies. The competitors include G-EDF, G-NP-EDF, G-FIFO, gMUA, P-EDF and P-DASA. Our study reveals that the GUA class of algorithms accrue higher utility and satisfy greater number of deadlines than the deadline-based scheduling algorithms by as much as 750% and 600%, respectively. In addition, ...

An experimental evaluation of the scalability of real-time scheduling algorithms on large-scale multicore platforms

“…Clark [40] and Welch [108] describe two distributed radar tracking systems which often must operate in overload, and suffer from non-deterministic execution times due to data-dependent processing and communication latencies incurred by their distributed structure. Both of their applications desire graceful performance degradation, and so require utility accrual scheduling [52], [39]. Tan and Hsu [103] also present a class of multimedia applications which may require guarantees on deadlines missed while this system is in overload.…”

“…However, aside from adding real-time scheduling capabilities, it makes no modifications to the Linux kernel to improve determinism or reduce latencies, and thus cannot be considered a real-time operating system. Furthermore, it provides no support for utility accrual scheduling of any kind [52].…”

“…This is a commonly made assumption when evaluating real-time systems, such as in [13], [52], [31]. However memory-intensive workloads are likely to scale differently than CPU-intensive workloads on large-scale core-count platforms because of the difference in migration and preemption overheads.…”

“…This is the most common model employed. Scheduling algorithms which use it include Global EDF (G-EDF) [24], LLREF [37], PFair [58], and the GUA class of algorithms [52].…”

“…The concept of a hard deadline is not implicit in TUFs, but can easily be expressed as a downward step function (see Figure 3.4(b)) [52]. Like Clark, Garyali, and others, in this thesis we consider only a downward step TUF in which a utility V i may be accrued if the job T j i is completed before its deadline d j i , and no utility is accrued if it is completed afterward [39], [52].…”