Donald W. Gillies

Scheduling Tasks with AND/OR Precedence Constraints

In traditional precedence-constrained scheduling a task is ready to execute when all its predecessors are complete. We call such a task an AND task. In this paper we allow certain tasks to be ready when just one of their predecessors is complete. These tasks are known as OR tasks. We analyze the complexity of two types of real-time AND/OR task scheduling problems. In the first type of problem, all the predecessors of every OR task must eventually be completed, but in the second type of problem, some OR predecessors may be left unscheduled. We show that most problems involving tasks with individual deadlines are NP-complete, and then present two priority-driven heuristic algorithms to minimize completion time on a multiprocessor. These algorithms provide the same level of worst-case performance as some previous priority-driven algorithms for scheduling AND-only task systems.

User Experience Modeling for DASH Video

Ever since video compression techniques have been introduced, measurement of perceived video quality has been a non-trivial task. Recently, a new class of video transport techniques has been introduced for transmission of video over varying channels such as wireless network. These transport techniques, called adaptive streaming, vary the bit rate and quality of the transmitted video to match the available channel bandwidth. DASH, Dynamic Adaptive Streaming over HTTP, is a new worldwide standard for adaptive streaming of video, audio and other media such as closed captioning. The adaptive streaming techniques introduce an additional level of complexity for measuring perceived video quality, as it varies the video bit rate and quality. In this paper, we study the perceived video quality using DASH. We investigate three factors which impact user perceived video quality: initial delay, stall (frame freezing), and bit rate (frame quality) fluctuation. Moreover, for each factor, we explore multiple dimensions that can have different effects on perceived quality. For example, in the case of the factor stall, while most previous research have studied how stall duration correlates with user perceived quality, we also consider when the stalls happen and how the stalls are distributed, since we believe they may also impact user experience. We design and conduct extensive subjective tests to study the impairments of the different dimensions of the three factors on user perceived video quality. We will describe the methodology to design the subjective tests, and present the results of the subjective tests. Based on the subjective tests, we derive impairment functions which can quantitatively measure the impairment of each factor on the user experience of any DASH video, and also provide validation results.

Scheduling tasks with AND/OR precedence constraints

In traditional precedence-constrained scheduling a task is ready to execute when all its predecessors are complete. Such a task is called an AND task. The authors allow certain tasks, known as OR tasks, to be ready when just one of their predecessors is complete. They analyze the complexity of two types of real-time AND/OR task scheduling problems. In the first type of problem, all predecessors of every OR task must eventually be completed, but in the second type of problem, some OR predecessors may be left unscheduled. The authors present two priority-driven heuristic algorithms that may be used to schedule AND/OR task systems on m processors to minimize completion time, and analyze the worst-case performance of these algorithms.<<ETX>>

Scheduling imprecise computations to minimize total error

Abstract We consider the problem of scheduling tasks each of which is logically decomposed into a mandatory part and an optional part. The mandatory part must be executed to completion in order to produce an acceptable result. The optional part refines the result in order to reduce the error in the result. If the available processor time is insufficient, the optional part can be left unfinished. The error in the result produced by an aperiodic task is equal to the processing time of the unexecuted portion of the optional part. Our objective is to schedule the tasks so that the total error over all tasks is minimized. A preemptive algorithm for scheduling n aperiodic dependent tasks with rational ready times, deadlines, and processing times on a uniprocessor system is described. This algorithm is optimal in the sense that it finds feasible schedules meeting the time constraints and minimizing the total error of all tasks, whenever feasible schedules exist. The complexity of this algorithm is O(n2log2n). We also consider the problem of scheduling periodic tasks. For these tasks the error in the result of each task is measured in terms of the average error in the results produced over several consecutive periods. A class of preemptive, priority-driven algorithms can be used to find feasible schedules with a small average error. These algorithms are described and evaluated.

Recent Results in Real-Time Scheduling

This paper gives an overview of several recent results on real-time scheduling. Specifically, it presents the workload models that characterize the following types of tasks: tasks with AND/OR precedence constraints, tasks with temporal distance constraints, distributed tasks with overall deadlines, and tasks with deferrable deadlines. Suitable algorithms for scheduling tasks of each type are described.

An application- and management-based approach to ATM scheduling

Two frame‐based ATM scheduling strategies, Delayed Frame Queueing (DFQ) and Virtual Frame Queueing (VFQ), are described for real‐time network applications. Both strategies guarantee explicit upper bounds on delay and jitter on a per virtual connection (VC) basis without relying on per‐VC queueing. They also resolve the disadvantages often associated with other frame‐based schedulers. The DFQ scheduler employs Resource Management cells to enable work‐conserving forwarding at intermediate nodes while retaining the option for non‐work‐conserving forwarding at terminal nodes. This allows delay bounds to be uncoupled from jitter bounds. The VFQ scheduler requires no such overhead cell transmissions, but is more limited in its ability to uncouple these two types of bounds. ATM network simulation results indicate that both proposed disciplines offer favorable multiplexing performance in comparison to a well documented high performance service discipline. Finally, we contrast our proposal against the significant difficulties that alternative scheduling strategies based on bandwidth guarantees will face with regard to network management and network synthesis issues.

An efficient ATM voice service with flexible jitter and delay guarantees

A transmission and multiplexing strategy appropriate for voice over asynchronous transfer mode (ATM), called delayed frame queueing (DFQ), is proposed. This frame-based strategy has features in common with the synchronous transfer mode and is thus well suited to service synchronous applications such as voice, while retaining the statistical multiplexing capabilities of ATM. In particular, the DFQ service discipline can provide explicit and nontrivial bounds for queue delay and jitter, for both bursty as well as continuous traffic streams. Furthermore, the DFQ discipline can combine a wide range of delay and jitter bounds while also managing the distribution of quality of service violations among the traffic streams when congestion occurs. Jitter control is performed at the network periphery and thus does not negatively influence multiplexing gain at intermediate nodes. This efficient strategy has major implications in terms of the preferred alternatives chosen by clients when implementing source clock recovery for voice. DFQ allows the entire range of implementation alternatives for voice over ATM to be appropriately serviced, such as ATM adaptation layer types 1 and 2 (AAL1/2), adaptive playout, and immediate playout.

Evaluation of statistical multiplexing performance for real-time ATM traffic employing a frame-based management strategy

This paper presents ATM network simulation results which explore the statistical multiplexing gain characteristics of virtual frame queueing (VFQ), a frame-based congestion management strategy with independent cell delay and cell delay variation bounds. The multiplexing performance is investigated using bursty sources requiring strict delay bounds, and is compared to the multiplexing performance offered by a work-conserving strategy called self clocked fair queueing. The simulation results suggest that the VFQ strategy is very efficient at high link utilization levels. Both strategies display similar multiplexing performance at lower link utilization levels. The discussion explores a variety of circuit-level and network-level issues that determine desirable qualities for ATM congestion management strategies.

Greed in resource scheduling

An examination is made of the worst-case performance of a class of scheduling algorithms commonly known as priority-driven or list-scheduling algorithms. These algorithms have anomalous, unpredictable performance when used to suboptimally schedule nonpreemptive tasks with precedence constraints. A general method is presented for deriving the worst-case performance of these algorithms. This method is easy to use, yet powerful enough to yield tight performance bounds for many classes of scheduling problems. It is demonstrated that the method has wide applicability. Several task systems are presented for which list-scheduling algorithms have worst-case performance, and the general characteristics of these task systems are discussed. It is believed that these task systems are sometimes ignored in simulation studies; consequently, the results of these studies may be overly optimistic.<<ETX>>

A distributed protocol to schedule and monitor ATM real-time traffic

A service discipline for real-time ATM traffic is presented. End-to-end delay and delay variation bounds are provided by transmitting cells within link-level frames. The service discipline does not require per VC queueing, and thus can be implemented using a fast and simple FIFO architecture. Finally, the service discipline is distributed, thus conformance to the end-to-end delay/jitter bounds can be monitored and controlled on a per node basis.