集群环境下I/O文件传输调度器的设计与实现

近年来研究人员对高性能计算中的并行I/O问题进行了深入研究,然而这些研究主要针对MPP问题,而对集群计算机系统中并行I/O问题的研究不多。因此,对于集群计算中的并行I/O系统进行研究是一个重要的研究课题。对集群计算中的并行I/O传输调度效率进行研究,设计了一个文件传输调度器,可以实现文件传输最快捷,节点资源最大利用,显著提高I/O节点吞吐率和反应时间。经过大量数据的测试和实验证明该调度器的有效性和适用性。     

一种基于分布式调度机制的集群体系结构

在分析现有集群器调度机制的基础上提出了一种基于分布式调度机制的集群体系结构，并与传统调度机制下的集群系统进行了对比分析。     

一种分布式调度中间件的设计和实现

本文提出了一种轻量级的用于大规模集群的分布式调度中间件产品方案,并结合业务需求,从应用架构、逻辑功能架构和总体技术架构等方面进行了详细的阐述。该方案对于研究和设计大中型互联网集群分布式调度软件有一定的参考借鉴意义。     

支持多优先级分组交换调度算法研究及其调度器设计

输入缓存交换结构的特点是缓存器和交换结构的运行速率与端口速率相等、实现容易,但存在队头阻塞。如果采用虚拟输出排队方法和适当的分组调度算法可予以消除,使吞吐率达到100%。文章首先研究讨论了并行迭代匹配算法,滑动迭代匹配调度算法的基本原理、迭代仲裁步骤及其硬件实现;对高速分组交换调度算法的性能进行了分析比较。然后给出了在高速输入队列交换机中实现多优先级调度算法的调度器设计与实现方案。经设计实现证明高速分组交换调度算法不仅硬件实现简单,而且具有良好的特性。     

Linux集群分布式安全研究

现有安全机制不能满足Linux集群的需要.本文通过对Linux集群的分布式安全进行研究,以进程作为基本安全实体,将强制访问控制机制由结点级扩展到整个集群,提出了一个在集群内实现安全区的分布式安全模型,用LMBehch3.0进行了测试,给出了性能比较结果.     

浅析LVS集群的负载均衡调度算法

采用LVS集群技术搭建所需的服务器,各服务器系统资源利用率经常存在很大差距。要解决该问题,必须依赖有效的负载均衡算法。该文深入研究LVS集群中常用的4种负载均衡算法,分析常用4种算法的优缺点,为搭建所需的服务器集群提供依据。     

多版本调度器的分析

本文针对目前许多多版本调度器算法在实际运行中性能不佳的情况,从理论上研究了多版本调度器的能力,并在此基础上提出了改善多版本调度器性能的方法。文中指出:1)多版本调度器可以对事务中的只读和只写操作做特殊处理;2)实际的多版本调度器除了必须满足可串行性原则外还应该满足合理性原则;3)采用保守方式对多版本调度器具有特别重要的意义。     

分布式Kahn处理网络的一种集群调度算法*

研究并提出一种采用分布式Kahn处理网络表达的并行程序在多处理器集群环境下的任务——处理器动态分配算法。由于Kahn处理网络的不可判定性,静态作业调度算法不能适用,而忽略其显式数据依赖关系的动态负载均衡策略存在很大的随机性,往往带来不必要的进程迁移。基于运行时动态生成的离散事件序列,预测Kahn处理网络在不同分配方案下的执行效率（处理器资源利用率）,迭代寻求最优动态分配方案,仿真效果良好。     

基于Linux集群的分布式实时操作系统的设计

文章针对Linux集群环境,以Linux通用操作系统为基础,提出了一种分布式实时操作系统的设计思路。以分布式实时操作系统的特殊性为线索,详细阐述了分布式实时操作系统中实时任务调度机制、实时进程间通信机制和实时网络通信机制的设计思路。     

多集群作业调度系统MCJSS的设计与实现

研究多集群的管理模式,设计支持多集群的用户决策和系统决策两种调度方式,提出基于PBS的多集群作业调度MCJSS架构。利用PBS的扩展接口,实现MCJSS的核心模块,管理多集群作业提交、作业转发和负载信息收集。给出了多集群间基于预测最轻负载转发的调度策略。运用两层负载信息收集策略服务于转发机制,利用组合单项负载的阈值来判断转发时机和目的集群。实验证明,相同的作业在相同规模但是不同组织形式的多集群和单集群组织模式下,多集群的系统吞吐量大于单集群组织模式。     

A general distributed scalable grid scheduler for independent tasks

We consider non-preemptively scheduling a bag of independent mixed tasks (hard, firm and soft) in computational grids. Based upon task type, we construct a novel generalized distributed scheduler (GDS) for scheduling tasks with different priorities and deadlines. GDS is scalable and does not require knowledge of the global state of the system. It is composed of several phases: a multiple attribute ranking phase, a shuffling phase, and a task-resource matched peer to peer dispatching phase. Results of exhaustive simulation demonstrate that with respect to the number of high-priority tasks meeting deadlines, GDS outperforms existing approaches by 10%–25% without degrading schedulability of other tasks. Indeed, with respect to the total number of schedulable tasks meeting deadlines, GDS is slightly better. Thus, GDS not only maximizes the number of mission-critical tasks meeting deadlines, but it does so without degrading the overall performance. The results have been further confirmed by examining each component phase of GDS. Given that fully known global information is time intensive to obtain, the performance of GDS is significant. GDS is highly scalable both in terms of processors and number of tasks—indeed it provides superior performance over existing algorithms as the number of tasks increase. Also, GDS incorporates a shuffle phase that moves hard tasks ahead improving their temporal fault tolerance. Furthermore, since GDS can handle mixed task types, it paves the way to open the grid to make it amenable for commercialization. The complexity of GDS is O(n²m)$O\left(n^{2}m\right)$ where n$n$ is the number of tasks and m$m$ the number of machines.     

A cautious scheduler for multistep transactions

Given a classC of serializable schedules, a cautiousC-scheduler is an on-line transaction scheduler that outputs schedules in classC and never resorts to rollbacks. Such a scheduler grants the current request if and only if the partial schedule it has granted so far, followed by the current request, can be extended to a schedule inC. A suitable extension is searched among the set of all possible sequences of the pending steps, which are predeclared by the transactions whose first requests have already arrived. If the partial schedule cannot be extended to a schedule inC, then the current request is delayed. An efficient cautiousCPSR-scheduler has been proposed by Casanova and Bernstein.This paper discusses cautiousWRW-scheduling, whereWRW is the largest polynomially recognizable subclass of serializable schedules currently known. Since cautiousWRW-scheduling is, in general, NP-complete as shown in this paper, we introduce, a subclass (namedWRW ^#) ofWRW and discuss an efficient cautiousWRW ^#-scheduler. We also show that the fixed point set of the cautiousWRW ^#-scheduler properly containsCPSR. Therefore, ourWRW ^#-scheduler allows more concurrency than anyCPSR- scheduler.This work was supported in part by the Natural Sciences and Engineering Research Council of Canada under Grant No. A5240 and in part by the Ministry of Education, Science and Culture of Japan under Scientific Research Grant-in-Aid.     

基于概率预测的P2P分布式存储副本查找

分布式存储系统是Peer-to-Peer技术的一个重要研究领域。如何实现有效的副本存储和副本查找定位是分布式存储系统的关键。DHT（Distributed Hash Tables）能够有效地实现分布式存储。在基于DHT的Peer-to-Peer分布式存储系统中提出了一种基于概率预测的副本查找和定位算法。分析并且详细描述了该算法,实现了算法的仿真模拟,分析了其性能。模拟结果显示,基于概率预测的副本查找算法通过概率的先预测能够实现高效率及高查准率的副本查找。     

实时系统优先级位图调度算法的改进

实时操作系统要求具有速度快和可预测性的特点,必须保证实时任务在要求的时间内完成。本文在描述了优先级位图调度算法的机制后,然后针对其在支持任务数量,所需数据空间和存在多个相同优先级任务的处理方面,提出改进方法。     

基于Agent分布式仪表产品造型协同设计应用研究

基于Agent代理,提出将Web技术和Multi-Agent技术融合的分布式协同设计系统,为产品设计过程中设计人员之间的信息共享和资源共享提供了更有效的方法.在分析系统体系结构的基础上,着重阐述了Multi-Agent代理模型和工作原理,并将其应用于仪表产品造型设计系统中,实现基于Agent的分布式仪表产品造型设计原型系统.     

基于结构化对等网络的高效分布式任务调度

提出与描述了一个基于结构化对等网络的高效分布式任务调度策略HDTS(High-Efficient Distributed Task Scheduling)。HDTS建立在结构化对等网络的应用层覆盖网络上,保证了系统的非集中性、可扩展性、自组织性,以及规模大的优点。HDTS把基于Chord风格的对等网络协议和容错及高效的多播调度策略结合起来,允许分布式计算的各种大量的子任务在对等网络的节点上高效的调度、分配、执行。除了支持主-从风格的并行计算外,系统允许具有数据依赖的分布式旅行商算法正确的执行,使系统具有通用性和开放性。测试结果表明:HDTS具有正确性与高效性,可以作为对等网络上计算资源高效共享与聚集的可行方案。     

分布协同综合系统体系结构的研究与设计

在分布式协同实验中,确定协同系统的体系结构是解决系统诸多关键技术的首要工作。在综合了其它协同系统体系结构的优点之上,提出了数据层、中间层和用户层三层体系结构,着重论述了中间层的设计和实现方法,并通过具体应用,实现了分布式实验环境下协同综合系统的基本功能,并验证了该体系结构的合理性和先进性。     

A distributed scheduler for air traffic flow management

A system was developed to efficiently schedule aircraft into congested resources over long ranges and present that schedule as a decision support system. The scheduling system consists of a distributed network of independent schedulers, loosely coupled by sharing capacity information. This loose coupling insulates the schedules from uncertainty in long-distance estimations of arrival times, while allowing precise short-term schedules to be constructed. This ??rate profile?? mechanism allows feasible schedules to be produced over long ranges, essentially constructing precise short-range schedules that also ensure that future scheduling problems are solvable while meeting operational constraints. The system was tested operationally and demonstrated reduced airborne delay and improved coordination.     

P2P-FISM: Mining (recently) frequent item sets from distributed data streams over P2P network

Data intensive large-scale distributed systems like peer-to-peer (P2P) networks are finding large number of applications for social networking, file sharing networks, etc. Global data mining in such P2P environments may be very costly due to the high scale and the asynchronous nature of the P2P networks. The cost further increases in the distributed data stream scenario where peers receive continuous sequence of transactions rapidly. In this paper, we develop an efficient local algorithm, P2P-FISM, for discovering of the network-wide recent frequent itemsets. The algorithm works in a completely asynchronous manner, imposes low communication overhead, a necessity for scalability, transparently tolerates network topology changes, and quickly adapts to changes in the data stream. The paper demonstrates experimental results to corroborate the theoretical claims.