Jian (Denny) Lin

Maximizing Guaranteed QoS in (m, k)-firm Real-time Systems

(m, k)-firm constraints have been used to schedule tasks in soft/firm real-time systems under overloaded conditions. In general, they are provided by application designers to guarantee the minimum levels of quality of service (QoS). Many problems concentrating in task schedulability under these constraints were investigated. However, little work has been done in combining the optimization of the QoS and task schedulability subject to these (m, k)-firm constraints. In this paper, we consider the problem of maximizing the guaranteed performance while maintaining a schedulable task set in periodic firm real-time systems. To quantify the performance, we propose a granularity-related metric called granularity of quality of service-reward (GQoS-reward). We then show that maximizing the total GQoS-reward is an NP-hard problem and a heuristic method to solve the problem is studied. In addition to the improvement to the GQoS, positive effects on other main performance metrics for soft/firm real time systems, such as effective processor utilization (EPU), total accumulated reward and instability, are also supported by the simulation results using our optimization strategy

Real-energy: a new framework and a case study to evaluate power-aware real-time scheduling algorithms

In the past decades, many algorithms with the goal of achieving energy efficiency have been proposed for scheduling real-time tasks. Due to a lack of a unified testing framework, most of them were evaluated via simulations under their own experimental scenarios. However, finding their performance in real processors is essential if these algorithms are to be used in practice. In this paper, we design a unified framework to evaluate power-aware scheduling algorithms based on a real Intel PXA255 XScale processor, and present a case study to compare several key algorithms using DVS/Shut-Down. The energy efficiency and the quantitative difference in their performance as well as the practical issues found in the implementation of these algorithms are discussed. Our experiments show a gap between the theoretical results and the real results. Our framework not only gives researchers a tool to evaluate their system designs, but also helps them to bridge this gap in their future works.

Real-Time Task Assignment in Heterogeneous Distributed Systems with Rechargeable Batteries

Real-time systems are one of the fields of computing where major benefits are expected from the increasing availability of multiprocessor technology. Heterogeneous computing environments, which utilize different high-performance machines interconnected via a high speed communication system, are well suited to the large, computation intensive, real-time or non-real-time applications. Nowadays, many of the systems in these environments are powered by rechargeable batteries. Scheduling real-time tasks on these rechargeable systems is an important issue which has been studied in the literatures. In this paper, we explore the task assignment problem on heterogeneous distributed system with rechargeable batteries. Our techniques to solve the problem are based on four heuristics, namely Minimum Schedule Length (MSL), min-min schedule length (MmSL), genetic algorithm (GA), and ant colony optimization (ACO). While the modifications of the MSL, MmSL and GA approaches from their original implementation are somewhat straight-forward, we design a novel structure using ACO. The performance comparisons of these four techniques are performed and the results are discussed. This paper not only gives a suggestion on which heuristic is best suited for the specific problem, but also provides a new direction to solve similar problems.

Real-Time Task Assignment in Rechargeable Multiprocessor Systems

This paper introduces the scheduling of frame-based real-time tasks in partitioning schemes for multiprocessor systems powered by rechargeable batteries. In frame-based real-time systems, a set of tasks must execute in a frame, and the whole frame is repeated. This system model is widely used in real-time communication, real-time imaging and a lot of other real-time/embedded systems. Nowadays, many of these systems are powered by rechargeable batteries. Scheduling real-time tasks on these rechargeable systems is an important yet largely ignored issue. The problem for uniprocessor systems had been studied in (A. Allavena and D. Mosse, 2001), in which an algorithm of complexity O(N) was proposed for determining the feasibility of the task set. However, it poses a challenge when doing so in a rechargeable multiprocessor system considering different characteristics of the batteries. In this paper, we first show this problem to be NP-hard, and then propose efficient algorithms to overcome it. The simulation results have shown that our algorithms exhibit very good behaviors and they can be considered as solutions to the problem.

Energy reduction for scheduling a set of multiple feasible interval jobs

Time-critical jobs in many real-time applications have more than one feasible interval. Such jobs can be executed in any of their feasible intervals. Given a schedulable set of multiple feasible interval (MFI) jobs, energy can be saved by carefully selecting the executing interval for each job. In this paper, we explore the energy minimization problem for real-time systems in which jobs have more than one feasible interval. The static and dynamic energy management schemes are both investigated to minimize the energy consumption while preserving the systems feasibility. Focusing on EDF scheduling algorithm, we first study reducing the dynamic power consumption of a single CPU. We show that the static optimal speed assignment problem is NP-Hard and propose a simulated annealing (SA) based approach to solve it. Then, we develop several on-line greedy algorithms to exploit run-time slack by reselecting a jobs executing interval on-the-fly. In addition, a leakage-aware version is discussed to improve the overall energy efficiency. Simulation results show that all proposed schemes achieve significant improvements of energy efficiency while the system remains schedulable.

Real-time Task Assignment with Replication on Multiprocessor Platforms

Fault tolerance is a very important aspect in critical real-time task scheduling. On multiprocessor systems, executing tasks with replication provides an additional reliability to resist potential processor failures and computing faults. For assigning real-time tasks on such systems, there must be requirements that all tasks assigned on the system meet their timing constraints, and all replicas of the same task are assigned to distinct processors. Obviously, such a reliability requirement could overload the system. In these situations, how to assign the tasks on processors to achieve the highest benefit poses a challenge. In this paper, we consider the problem of maximizing the number of successfully assigned tasks on a homogeneous distributed multiprocessor system, while satisfying the real-time constraint and system reliability requirement. Exact, greedy approximation and polynomial time approximation scheme (PTAS) algorithms are developed for the problem. Theoretical analysis, necessary proofs and experimental results that support our claims are all given.

Power-Aware Scheduling for Multiple Feasible Interval Jobs

Time-critical jobs in many real-time applications have more than one feasible interval. Such jobs can be executed in any of their feasible intervals. Given a Multiple Feasible Interval (MFI) job set that is schedulable, energy can be saved by carefully selecting the executing interval for each job. In this paper, we explore the energy minimization problem for real-time systems in which jobs have multiple feasible intervals. The static and dynamic energy management schemes are both investigated to minimize the energy consumption while preserving the systempsilas feasibility. Focusing on the EDF scheduling algorithm, we first study reducing the dynamic power consumption. We show that the static optimal speed assignment problem is NP-Hard and propose a Simulated Annealing (SA) based approach to solve it. Then, we develop an online greedy algorithm to exploit the run-time slacks by ldquofetchingrdquo the eligible job from a hot spot to execute earlier, thus, reducing the dynamic energy consumption. In addition, a leakage-aware version is discussed to improve the overall energy efficiency as well. Simulation results show that all the proposed schemes can achieve significant improvements on energy efficiency while the system remains schedulable.

Approximation algorithms in partitioning real-time tasks with replications

Abstract Today is an era where multiprocessor technology plays a major role in designs of modern computer architecture. While multiprocessor systems offer extra computing power, it also opens a new range of opportunities to improve fault-robustness. This paper focuses on a problem of achieving fault-tolerance using replications in real-time, multiprocessor systems. In the problem, multiple replicas, or copies, of a computing task are executed on distinct processors to resist potential processor failures and computing faults. Two greedy, approximation heuristics, named Worst Fit Increasing K-Replication and First Fit Increasing K-Replication, are studied to maximise the number of real-time tasks assigned on a system with identical processors, respecting to the tasks’ replicating and timely requirements. Worst case performance is analysed by using an approximation ratio between the algorithms and an optimal solution. We mathematically prove that the ratios of using both algorithms are infinitely close to 2. Simulations are performed on a large set of testing cases which can be used to bring to light the average performance of using the algorithms in practice. The results show that both heuristic algorithms provide simple but fast and effective solutions to solve the problem. Graphical Abstract Assigning real-time tasks to a multiprocessor system with replications.

Scheduling Mixed-Criticality Real-Time Tasks in a Fault-Tolerant System

Enabling computer tasks with different levels of criticality running on a common hardware platform has been an increasingly important trend in the design of real-time and embedded systems. On such systems, a real-time task may exhibit different WCETs (Worst Case Execution Times) in different criticality modes. It is well-known that traditional real-time scheduling methods are not applicable to ensure the timely requirement of the mixed-criticality tasks. In this paper, the authors study a problem of scheduling real-time, mixed-criticality tasks with fault tolerance. An optimal, off-line algorithm is designed to guarantee the most tasks completing successfully when the system runs into the high-criticality mode. A formal proof of the optimality is given. Also, a novel on-line slack-reclaiming algorithm is proposed to recover from computing faults before the tasks’ deadline during the run-time. Simulations show that an improvement of about 30% in performance is obtained by using the slack-reclaiming method. Scheduling Mixed-Criticality Real-Time Tasks in a Fault-Tolerant System

Lossless Pruned Naive Bayes for Big Data Classifications

Abstract In a fast growing big data era, volume and varieties of data processed in Internet applications drastically increase. Real-world search engines commonly use text classifiers with thousands of classes to improve relevance or data quality. These large scale classification problems lead to severe runtime performance challenges, so practitioners often resort to fast approximation techniques. However, the increase in classification speed comes at a cost, as approximations are lossy, mis-assigning classes relative to the original reference classification algorithm. To address this problem, we introduce a Lossless Pruned Naive Bayes (LPNB) classification algorithm tailored to real-world, big data applications with thousands of classes. LPNB achieves significant speed-ups by drawing on Information Retrieval (IR) techniques for efficient posting list traversal and pruning. We show empirically that LPNB can classify text up to eleven times faster than standard Naive Bayes on a real-world data set with 7205 classes, with larger gains extrapolated for larger taxonomies. In practice, the achieved acceleration is significant as it can greatly cut required computation time. In addition, it is lossless: the output is identical to standard Naive Bayes, in contrast to extant techniques such as hierarchical classification. The acceleration does not rely on the taxonomy structure, and it can be used for both hierarchical and flat taxonomies.