随着分布式系统的广泛应用，提高系统的性能的动态负载平衡策略也日益为人们所视。本文提出了中间应用服务器负载的一种动态均衡策略，并在Java RMI技术的基础上构建了使用该策略的均衡模型，最后给出了该模型相应的实现算法。

本文对网络模拟中必需的自相似通信量生成方法进行了研究，提出了一种基于分数高斯噪声（FGN, fractional Gaussian noise）过程和逆离散傅立叶变换（IDFT, inverse discrete Fourier transform）的自相似通信量生成算法——FGN-IDFT算法，并通过仿真实验和比较分析的方法对该算法的精度和计算复杂度进行了大量的实验和分析。研究结果表明，利用FGN-IDFT算法生成的自相似通信量的精度和速度都令人满意，为实际网络模拟的进一步深入研究提供了基础。

Web服务资源地有效组织与获取是网格计算中的关键性问题。本文基于OGSA（open grid service architecture）框架提出了一种有效地网格Web服务资源组织体系模型，它将Web服务资源的逻辑结构与物理资源组织分开，以层次结构的Web服务解析系统（Web Services Name System, WSNS）来进行Web服务资源查找与定位，以区域自治系统（Area Autonomy System, AAS）来进行Web服务资源的维护。并详细描述了该模型的实现机制、组织协议与数据结构。通过对模型的分析和对原型系统的模拟表明该模型是合理而有效的：它可以在较小开销下获得较满意的资源组织与定位性能，能适应网格Web服务资源的异构、复杂与动态性，多级的负载平衡，良好的可扩展性。

Critical Web applications, such as E-Commerce etc, require relatively high differential service ensured. However, current differential service control algorithms seldom take into account the network access characteristics. In this paper, a new admission control algorithm considering the network self-similar access characteristics is proposed. Taking advantage of the mathematical model of the network traffic admission control which can effectively overcome the self-similar characteristics of the network requests, through the scheduling of the differential service queue based on priority while at the same time taking into account various factors including access characteristics of requests, load information, etc, smoothness of the admission control is ensured by the algorithm proposed in this paper. We design a non-linear self-adapting control algorithm by introducing an exponential admission function, thus overcomes the negative aspects introduced by static threshold parameters. Simulation results show that the scheme proposed in this paper can effectively improve the resource utilization of the clusters, while at the same time protecting the service with high priority. Our simulation results also show that this algorithm can improve system stability and reliability too.

本文针对目前电子商务等关键Web用的软实时容错需求。提出了一种基于概率的容错调度算法，算法以概率形式进行允可接纳控制，克服了以往的调度算法没有考虑任务的自相似访问特征的不足，同时用概率对任务的可调度性进行分析。算法采用了任务的主/从备份技术来进行容错，当主任务无错时，回收从任务执行资源。实验结果表明该算法能较好地控制可接纳区域，有效地提高了系统资源的利用率。

In this paper, we discuss how to apply the LPT algorithm to find the database server number and the middle-tier application server number in order to reduce the costs.

A new static task scheduling algorithm named edge-zeroing based on dynamic critical paths is proposed. I he main ideas of the algorithm are as follows; firstly suppose that all of the tasks are in different clusters; secondly, select one of the critical paths of the partially clustered directed acyclic graph; thirdly, try to zero one of graph communication edges; fourthly. repeat above three processes-until all edges are zeroed; finally, check the generated clusters to see if some of them can be further merged without increasing the parallel time. Comparisons of the previ ous algorithms with edge-zeroing based on dynamic critical paths show that the new algorithm has not only a low complexity but also a desired performance comparable or even better on average to much higher complexity heuristic algorithms.

Abstract. Based on analysis on multiple packet losses of standard slow start caused by exponential growth of congestion window (cwnd), this paper proposes a new phase-divided TCP start scheme and designs a parameterized model to reduce packet losses and improve TCP performance. This scheme employs different of cwnd growth rules while cwnd is under and over the value of half window threshold (ssthresh) respectively, namely exponential growth and negatively exponential growth, which greatly decreases probability of multiple packet losses from a window of data and guarantees that a connection smoothly joins the Internet and transforms into congestion avoidance. Parameterized model adjusts the duration of slow start and acceleration of increasing cwnd to improve performance of slow start phase through various parameter setting. An adaptive paraeter setting method is designed. And the simulation results show that this new method significantly decreaseds packet losses and improves the stability of TCP and performance of slow start, and also achieves good fairness and friendliness to other TCP connections.

Aiming at the peculiarities of the resource organization in distributed, heterogeneous grid environment, a new protocol based on resource tree is proposed in this paper. The resource tree is constructed through bounded two-level broadcasting technique. We investigate the data structures used by this protocol and the resource tree construction algorithm in detail in this paper. Our simulation results show that the protocol proposed in this paper has a number of advantages, including: low node storage overhead, supporting dynamic resource addition & deletion, effectively hiding the heterogeneity of the resources in wide-area networks, high scalability of resource organization, fast construction of resource trees and high resource searching performance, etc. Although there is some broadcasting overhead involved, we can still get relative high performance due to the limited number of control packets required in this protocol.

For the parallel tasks represented by the Directed Acyclic Graph (DAG), if it is linearly clustered, the ordering of the execution time of the tasks in each cluster is based on their arrows in the DAG. But for nonlinearly clustering, the ordering of the independent tasks in each cluster is not easily decided. Improper ordering of these independent tasks will greatly increase the scheduling length of the DAG. We discuss the shortcomings of current scheduling algorithms and the reason behind poor performance, and then propose some new node information to be extracted which is used by a new independent tasks scheduling algorithm based on the Maximized Parallelism Degree (MPD). Experimental results show that the MPD algorithm can yield better performance than the previous algorithms.