面向多用户环境的MapReduce集群调度算法研究

针对MapReduce集群现有调度策略在多用户环境下无法根据用户的实际资源需求实现动态资源分配的问题,提出了一种基于历史执行信息(HEI)的MapReduce集群调度算法——HEI Scheduler。该算法通过建立集群作业执行信息的收集和分析机制,得到各用户组资源需求随时间变化的规律,并以作业实际占用slot的时间作为作业占用资源量的衡量标准,进而动态地确定资源池的最小共享资源以及集群剩余资源分配的权值。实验结果表明,执行信息分析机制能够更准确地表征作业对资源的需求,采用集群调度算法HEI Scheduler能够有效地缩短作业的整体执行时间。     

带准备时间的异构并行机调度规则自动设计方法

以大规模定制化的家电行业生产为背景,将家电总装产线的投产排序决策抽象成为一类带准备时间的异构并行机动态调度问题。针对人工调度规则解决动态调度问题简单高效,但场景适应性弱的特点,引入了基于遗传规划 (genetic programming, GP) 的规则自动设计框架。首先,通过分析家电总装产线生产特征以及优化需求,以最小化平均拖期为优化目标,建立异构并行机调度模型;随后,针对问题特征,构建线体指派-工单排序规则对协同进化的改进型GP算法,并提取线体、工单的特征属性输入GP算法框架以自动设计调度规则。最后,基于某家电企业实际案例数据设计大量算例测试集,通过对比GP算法与人工设计规则在差异化工况场景的实验结果,验证GP算法有效性,并进一步分析了GP算法构造规则受不同生产环境参数的影响。     

异构负载下单机调度与自适应序贯视情维修联合决策

生产调度期间,加工对象差异会使设备承受不同载荷,退化率也会改变,即异构负载现象。若设置相同的预防性维修阈值可能会导致设备过维修或欠维修。因此,提出了联合调度与维修策略,采用自适应序贯视情维修,以期望完成时间最小为目标,构建性能可靠度约束下两者的联合决策模型。通过数值实验,对比考虑异构负载前后的优化结果,以验证考虑异构负载的必要性。联合决策时分别选取视情维修阈值和自适应序贯式的视情维修阈值序列进行优化,结果表明选取自适应序贯式的视情维修阈值序列优化得到的总加权期望完成时间可降低25.17%。对设备的性能可靠性阈值下界进行灵敏度分析,得出合理的性能可靠度约束范围,以确保设备运行安全性的同时尽量控制目标成本。     

K~+和Na~+在C_3S-纳米SiO_2浆体上的存在形式

以拌合液为KOH和NaOH混合溶液的C_3S-纳米SiO_2浆体为对象,研究了液固比对浆体在去离子水中释放K~+和Na~+的影响以及浆体对K~+和Na~+的滞留能力;并采用29Si MAS NMR分析讨论了K~+和Na~+在C_3S-纳米SiO_2浆体上的存在形式。结果表明:随着液固比的增大,浆体释放的K~+和Na~+的百分数逐渐增大;经去离子水充分洗涤后仍有一部分K~+和Na~+存在于浆体中,且浆体对K~+和Na~+的滞留能力随着纳米SiO_2掺量的增加以及拌合液碱浓度的增大而增强。29Si MAS NMR分析证实了K~+和Na~+能够进入C-S-H凝胶层间形成不可逆的化学结合。K~+和Na~+通过取代C-S-H凝胶层间≡SiOCa+中的Ca2+,缩短了C-S-H凝胶中[SiO4]4-四面体链长,从而使凝胶中Q1和Q2峰向左移动。     

气相传输平衡法在Sr2+:α-BBO衬底上制备β-BB0薄膜的研究

采用气相传输平衡技术在掺Sr2 的α-BBO(001)衬底上制备了β-BBO薄膜,研究了粉料配比、VTE处理温度以及恒温时间对薄膜质量的影响.结果表明,在适当的粉料配比、VTE处理温度以及恒温时间下,在Sr2 :α-BBO晶片表层制备呈c轴高度择优取向的、没有第二相物质的β-BBO薄膜,其双晶摇摆曲线半峰宽值FWHM为1000",显示出β-BBO薄膜较好的结晶质量.并对气相传输平衡技术制备β-BBO薄膜的机理进行了初步的探讨.     

在Sr 2+ :α-BBO单晶衬底上生长β-BBO薄膜的研究

采用液相外延法、脉冲激光沉积法以及气相传输平衡法在Sr^2＋：α-BBO（001）衬底上生长了质量优异的β-BBO薄膜,对薄膜进行了X射线衍射测试以及双晶摇摆曲线分析,结果表明,采用Sr^2＋：α-BBO单晶作为衬底制备的β-BBO薄膜具有高的择优取向度和低的半峰宽值,同目前制备β-BBO薄膜所采用的其他衬底材料相比,Sr^2＋：α-BBO和β—BBO之间具有结构相似、透光范围匹配以及化学稳定性匹配的优点,表明Sr^2＋：α-BBO单晶将是生长β-BBO薄膜的优异衬底材料．     

Er3+掺杂量对TiO2:Er粉体上转换发光性能的影响

通过溶胶-凝胶法制备了Er³⁺离子掺杂的TiO₂纳米粉体。通过X射线衍射、荧光检测对制备的纳米晶微结构、发光性能进行了表征。结果表明：纳米晶微结构及粒径随掺杂量呈规律性变化,掺杂稀土离子可抑制其相变。在980nm红外激光泵浦下,其发射峰分别位于527nm、548nm、660nm处,通过改变Er³⁺掺杂量,获得绿红强度比最高为29.3,最低为0.7,实现了单一掺杂改变发光颜色的调控。借助交叉弛豫与浓度猝灭理论,解释了其发光性能的规律性变化,并分析了其上转换机制。     

A fuzzy logic modelling of dynamic scheduling in FMS

This paper is concerned with scheduling in flexible manufacturing systems (FMSs) using a fuzzy logic (FL) approach. Four fuzzy input variables: machine allocated processing time, machine priority, due date priority and setup time priority are defined. The job priority is the output fuzzy variable, showing the priority status of a job to be selected for next operation on a machine. The model will first select the machines and then assign operations based on a multi-criteria scheduling scheme. The performance of the approach is compared against established methods reported in the literature. The performance measures considered average machine utilisation, meeting due dates, setup times, work in process and mean flow times. The test results demonstrate the superiority of the fuzzy logic approach in most performance measures.     

OFDMA系统中支持混合业务的跨层调度算法

针对正交频分多址接入(OFDMA)系统下行链路的混合业务调度问题,提出了一种基于队列等待时间的跨层调度算法.该算法联合利用了MAC层的队列等待时间与物理层的信道状态信息作为调度参数,通过队列等待时间反映用户的服务质量要求,并利用多用户分集增益提高系统性能;针对实时和非实时用户的不同服务质量要求,在队列等待时间的计算上采取了不同的策略;在子载波的分配过程中根据分配状态及时更新队列等待时间,使资源的利用更为有效.仿真结果表明,提出的算法可显著降低实时用户的平均时延和最大延时违反概率,同时保证了非实时用户的吞吐量需求,能够有效地支持下一代网络中混合业务的多种服务质量要求.     

DSS approach for heterogeneous parallel machines scheduling considering proximate supply chain constraints

This paper describes the basis of a Decision Support System (DSS) designed to schedule fertiliser production orders to be delivered within time windows, in plants made up of multiple heterogeneous parallel processors (production lines), considering that fertiliser production rates and nomenclatures depend on lines, that setup times depend on sequence and lines, and taking into account downtime constraints (preventive maintenance?…). A mixed linear programming model is encapsulated in the DSS which considers the schedule’s impacts, immediately upstream and downstream of plants in the supply chain. These side-effects may make the proposed solution unfeasible and the DSS helps redefining the problem to avoid them.     

Anticipation and flexibility in dynamic scheduling

Many real-world optimization problems change over time and require frequent re-optimization. We suggest that in such environments, an optimization algorithm should reflect the problem's dynamics and explicitly take into account that changes to the current solution are to be expected. We claim that this can be achieved by having the optimization algorithm search for solutions that are not only good, but also flexible, i.e. easily adjustable if necessary in the case of problem changes. For the example of a job-shop with jobs arriving non-deterministically over time, we demonstrate that avoiding early idle times increases flexibility, and thus that the incorporation of an early idle time penalty as secondary objective into the scheduling algorithm can greatly enhance the overall system performance.     

Multi-criteria optimisation-based dynamic scheduling for controlling FMS

This study deals with controlling flexible manufacturing systems (FMS) operating in volatile production environments. Most studies that address this issue use some sort of adaptive scheduling that enables the FMS to cope with the randomness and variability efficiently. The methods presented in the literature are usually based on heuristics and use simple dispatching rules. They do not consider changing the decision criteria dynamically as the system conditions change. In contrast to previous studies, the present study focuses on developing a control mechanism for dynamic scheduling that is based on incremental optimisation. This means that each time a scheduling decision is made, the local optimisation problem is solved such that the next jobs to be processed on machines are selected. The objective function (dominant decision criterion) for this optimisation problem is selected dynamically based on production order requirements, actual shop-floor status and system priorities. The proposed multi-criteria optimisation-based dynamic scheduling methodology was evaluated and compared with some known scheduling rules/policies. The results obtained demonstrate the superiority of the suggested methodology as well as its capability to cope with a multi-criteria environment.     

Self-adaptive dynamic scheduling of virtual production systems

To enhance the agility of virtual production systems (VPSs) under today's dynamic and changing manufacturing environment, a self-adaptive dynamic scheduling method based on event-driven is proposed for VPSs in this paper. This method is composed of the mechanisms and algorithm of self-adaptive dynamic scheduling. In the mechanisms, the dynamic events faced by VPSs are determined through users’ inputs or supervisory controllers’ detections, the local effects made on the schedule are analysed according to the dynamic events, and the self-adaptive measures and rules are specified correspondingly. To implement the dynamic scheduling of VPSs under the guidance of self-adaptive rules, a modified heuristic rescheduling algorithm is proposed for affected operations. A case study illustrates that the proposed method can well accomplish the dynamic scheduling of VPSs in a self-adaptive manner.     

基于聚类状态隶属度的动态调度Q-学习

提出了一种利用Q-学习解决动态单机调度环境下的自适应调度规则选择的方法.该方法针对动态调度环境中系统状态空间大,Q-学习不易收敛的特点,首先提取系统状态特征,对系统状态进行合理聚类,有效地降低系统状态空间维数,然后在学习过程中令设备Agent根据瞬时状态向量对各聚类状态的隶属度做出综合判断,选择合适规则,并在每次迭代后根据隶属度将动作奖惩分配给各聚类状态的动作值函数.仿真结果表明,所提Q-学习算法较之传统Q-学习具有更快的收敛速度,提高了设备Agent的动态调度规则选择能力.     

Performance evaluation of dynamic scheduling approaches in vehicle-based internal transport systems

This paper studies the performance of static and dynamic scheduling approaches in vehicle-based internal transport (VBIT) systems and is one of the first to systematically investigate under which circumstances, which scheduling method helps in improving performance. In practice, usually myopic dispatching heuristics are used, often using look-ahead information. We argue more advanced scheduling methods can help, depending on circumstances. We introduce three basic scheduling approaches (insertion, combined and column generation) for the static problem. We then extend these to a dynamic, real-time setting with rolling horizons. We propose two further real-time scheduling approaches: dynamic assignment with and without look-ahead. The performances of the above five scheduling approaches are compared with two of the best performing look-ahead dispatching rules known from the literature. The performance of the various approaches depends on the facility layout and work distribution. However, column generation, the combined heuristic, and the assignment approach with look-ahead consistently outperform dispatching rules. Column generation can require substantial calculation time but delivers very good performance if sufficient look-ahead information is available. For large scale systems, the combined heuristic and the dynamic assignment approach with look ahead are recommended and have acceptable calculation times.     

Intelligent dynamic control of stochastic economic lot scheduling by agent-based reinforcement learning

This paper addresses a stochastic economic lot scheduling problem (SELSP) for a single machine make-to-stock production system in which the demands and the processing times for N types of products are random. The sequence-independent setup times and costs are explicitly considered and may have different values for various types of products. The SELSP is to decide when, what, and how much (the lot size) to produce so that the long-run average total cost, including setup, holding and backorder costs, is minimised. We develop a mathematical model and propose two reinforcement learning (RL) algorithms for real-time decision-making, in which a decision agent is assigned to the machine and improves the accuracy of its action-selection decisions via a ‘learning’ process. Specifically, one is a Q-learning algorithm for a semi-Markov decision process (QLS) and another is a Q-learning algorithm with a learning-improvement heuristic (QLIH) to further improve the performance of QLS. We compare the performance of QLS and QLIH with a benchmarking Brownian policy and the first-come-first-served policy. The numerical results show that QLIH outperforms QLS and both benchmarking policies.