【摘要】：大規(guī)模和異構(gòu)性是現(xiàn)今網(wǎng)絡(luò)技術(shù)的快速發(fā)展方向,傳統(tǒng)的網(wǎng)絡(luò)體系結(jié)構(gòu)和資源分配調(diào)度策略等已經(jīng)沒有辦法滿足網(wǎng)絡(luò)的發(fā)展需要,尤其是在那些用來進(jìn)行大計(jì)算量任務(wù)的分布式系統(tǒng)中。其問題是在于大規(guī)模的網(wǎng)絡(luò)計(jì)算系統(tǒng)中,計(jì)算節(jié)點(diǎn)或存儲(chǔ)節(jié)點(diǎn)等呈現(xiàn)出分散的、分布的特點(diǎn),而且各種節(jié)點(diǎn)之間也是不盡相同的。這些節(jié)點(diǎn)的異構(gòu)性就導(dǎo)致了一些常用的分配算法很難適用,因此就需要多種不同的分配策略。除此之外,大規(guī)模網(wǎng)絡(luò)計(jì)算系統(tǒng)(網(wǎng)格計(jì)算、云計(jì)算)向商用化發(fā)展,它需要滿足的外部用戶的需求也更加的多樣化、復(fù)雜化,對(duì)系統(tǒng)提供服務(wù)的時(shí)效性、可靠性、安全性等的要求也越來越高;計(jì)算系統(tǒng)的服務(wù)商要在滿足用戶任務(wù)需求的情況下,降低自身運(yùn)營成本,這樣才能使得大規(guī)模網(wǎng)絡(luò)計(jì)算系統(tǒng)健康有效的長久發(fā)展下去。因此,對(duì)大規(guī)模計(jì)算系統(tǒng)中的多性能優(yōu)化、建模以及算法的研究是十分必要的。本文對(duì)于上述大規(guī)模網(wǎng)絡(luò)計(jì)算系統(tǒng)中出現(xiàn)的問題,給出了基于multi-agent的自主調(diào)度的算法,算法考慮了任務(wù)時(shí)效性和可靠性兩種性能要求,對(duì)任務(wù)進(jìn)行劃分并且在計(jì)算節(jié)點(diǎn)中分配。算法求解中使用到了通用生成函數(shù)和模擬退火算法,最后通過實(shí)驗(yàn)對(duì)模型進(jìn)行了驗(yàn)證。首先,在前人工作的基礎(chǔ)上進(jìn)行研究,根據(jù)大規(guī)模網(wǎng)絡(luò)計(jì)算系統(tǒng)和人工智能中多代理系統(tǒng)的機(jī)制,將兩者結(jié)合起來,構(gòu)造基于multi-agent的大規(guī)模自主計(jì)算系統(tǒng)模型。在建立的模型的基礎(chǔ)上,提出任務(wù)劃分和計(jì)算節(jié)點(diǎn)的分配算法。當(dāng)系統(tǒng)收到用戶的計(jì)算服務(wù)請(qǐng)求時(shí),本系統(tǒng)會(huì)根據(jù)服務(wù)的信息進(jìn)行分析,以確定任務(wù)是最短服務(wù)時(shí)間原則還是最低可靠性原則,并據(jù)此將之劃分成多個(gè)小的、可并行執(zhí)行的、耦合度低的EB(可執(zhí)行模塊),并將這些模塊分配給計(jì)算節(jié)點(diǎn)進(jìn)行執(zhí)行。其次,提出對(duì)系統(tǒng)各性能的評(píng)估數(shù)學(xué)模型,對(duì)任務(wù)執(zhí)行的時(shí)效性、可靠性進(jìn)行評(píng)估,然后再借助通用生成函數(shù)技術(shù)來對(duì)各個(gè)指標(biāo)進(jìn)行快速計(jì)算,提出基于最短服務(wù)時(shí)間和最低可靠性原則的服務(wù)計(jì)算模型。最后,針對(duì)本文中提出的優(yōu)化問題,采用模擬退火算法來對(duì)其進(jìn)行求解,并且實(shí)際仿真實(shí)驗(yàn)對(duì)其進(jìn)行驗(yàn)證并求得不同原則下的任務(wù)劃分和計(jì)算節(jié)點(diǎn)的分配,最終獲得問題的相對(duì)全局最優(yōu)解。
[Abstract]:Large-scale and heterogeneity are the rapid development direction of network technology nowadays. The traditional network architecture and resource allocation and scheduling strategy have been unable to meet the needs of network development. Especially in distributed systems that are used for large computational tasks. The problem lies in the scattered and distributed characteristics of computing nodes or storage nodes in large-scale network computing systems. Due to the heterogeneity of these nodes, some common allocation algorithms are difficult to apply, so many different allocation strategies are needed. In addition, large-scale network computing systems (grid computing, cloud computing) to commercial development, it needs to meet the needs of external users more diversified, complex, to provide services to the system timeliness, reliability, The requirement of security and so on is higher and higher. The service providers of computing system should reduce their own operating cost in order to make the large-scale network computing system develop healthily and effectively for a long time. Therefore, it is necessary to study multi-performance optimization, modeling and algorithms in large-scale computing systems. In this paper, an algorithm of autonomous scheduling based on multi-agent is presented to solve the problems in large-scale network computing systems. The algorithm takes into account the requirements of timeliness and reliability of tasks, divides the tasks and allocates them in computing nodes. The general generating function and simulated annealing algorithm are used in the solution of the algorithm. Finally, the model is verified by experiments. Firstly, based on the previous work, according to the mechanism of large-scale network computing system and multi-agent system in artificial intelligence, a large-scale autonomous computing system model based on multi-agent is constructed by combining the two mechanisms. On the basis of the established model, a task partition algorithm and an algorithm for computing node assignment are proposed. When the system receives the user's request for computing service, the system will analyze the information of the service to determine whether the task is the principle of minimum service time or the principle of minimum reliability, and divide it into several small, parallel execution. The EB (executable module) with low coupling degree is assigned to the computing node for execution. Secondly, a mathematical model for evaluating the performance of the system is proposed to evaluate the timeliness and reliability of task execution, and then to calculate each index quickly with the help of general-purpose generating function technology. A service computing model based on the principle of minimum service time and minimum reliability is proposed. Finally, in view of the optimization problem proposed in this paper, simulated annealing algorithm is used to solve the problem, and the actual simulation experiment verifies it and obtains the task partition and calculation node allocation under different principles. Finally, the relative global optimal solution of the problem is obtained.
【學(xué)位授予單位】：電子科技大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2014
【分類號(hào)】：TP393.02;TP301.6
，

本文編號(hào)：2151488