On the ineffectiveness of 1/m-based interference bounds in the analysis of global EDF and FIFO scheduling

Abstract

Enormous efforts have been spent in the derivation of sufficient schedulability tests for popular global schedulers such as global fixed-priority (G-FP) and global earliest-deadline first (G-EDF). Among all the proposals, response-time analysis techniques are established as the most popular ones and have been widely adopted by several authors in subsequent researches. Such tests are based on interference bounds that have been derived by exploiting the work-conserving property of such schedulers, and are typically expressed with a 1/m multiplier in the interference equation (with m being the number of available processors).

On the Trade-offs between Generalization and Specialization in Real-Time Systems

“…These analyses are, however, not able to show any gain compared to partitioned scheduling, even if task migration costs are ignored. On the contrary, it has been shown for global static-priority scheduling [66] as well as for global EDF and global FIFO scheduling [17] that the sufficient analysis for global scheduling is dominated by partitioned scheduling (i.e., all task sets that are deemed schedulable under global scheduling according to these tests are also schedulable under partitioned scheduling when a certain partitioning algorithm is used). Hence, a fundamentally different analysis technique is needed to exploit the potentially higher utilisation of global scheduling compared to partitioned scheduling.…”

Timing isolation and improved scheduling of deep neural networks for real‐time systems

“…Specifically, each TensorFlow thread is protected by a reservation server with budget and period chosen at design time. Furthermore, to leverage the benefits offered by partitioned scheduling, with respect to the pitfalls and the lower predictability provided by global schedulers, 40-42 each TensorFlow thread is statically allocated to one of the available cores. Unfortunately, due to the presence of precedence constraints in the parallel workload originated by TensorFlow, the timing protection of the TensorFlow threads via resource reservation can originate performance degradation: this problem is addressed in Section 5, where modifications to both the TensorFlow scheduler and SCHED_DEADLINE are presented.…”

Research on parameter optimisation of dynamic priority scheduling algorithm based on improved reinforcement learning

“…The study of a dynamic priority scheduling algorithm is limited to the effect of a single factor on deadline miss rate (DMR), hence real-time task scheduling has different task values, energy consumption [6,7], task deadlines [8], and task execution times. Real-time system scheduling research improves the success rate of task scheduling and reduces its DMR [9][10][11][12], task schedulability decision conditions [13], preemption threshold scheduling settings [14], and hierarchical framework scheduling [15].…”