تشخیص کیفی QoS آگاه از متن انتها به انتها و تضمین کمی بر ​​اساس شبکه های بیزی

To support Quality of Service (QoS) management on current Internet working with best effort, we propose a systematic approach for end-to-end QoS qualitative diagnosis and quantitative guarantee. Both QoS metrics and contexts of a service are considered in a comprehensive manner in our approach, which consists of three sequential stages: context discretization, QoS qualitative diagnosis and QoS quantitative guarantee. Based on Fuzzy set, an automatic unwatched discretization algorithm for discretizing continuous numeric-value is brought forth to reshape these QoS metrics and contexts into their discrete forms. For QoS qualitative diagnosis, causal relationships between a QoS metric and its contexts are exploited with the help of K2 Bayesian network (BN) structure learning by treating QoS metrics and contexts as BN nodes. A QoS metric node is qualitatively diagnosed to be causally related to its parent context nodes. An ordering method is proposed to arrange orders for nodes involved in K2 algorithm. To guarantee QoS quantitatively, those causal relationships are next modeled quantitatively by BN parameter learning. BN inference is referred to calculate the marginal on a QoS metric node given its tunable parent context nodes. Then, the QoS metric is guaranteed to a specific value a user demands with certain probability by tuning its causal contexts to suitable values suggested by BN inference, that is, QoS quantitative guarantee is reached by now. Simulations, on a peer-to-peer (P2P) network, about the above three sequential stages are discussed and our approach is validated to be soundable and effective. We also argue that our approach can be reached in a polynomial time complexity in practice.

Current Internet delivers services merely with its best efforts; Quality of Service (QoS) is only supported or guaranteed as much as possible. There are some reasons for this awkward situation. First, it is not easy to trace, locate and identify the causes of a QoS violation on the volatile Internet. Second, even if this problem is fixed up, other problems may arise, for example, for security consideration, integrated service providers (ISPs) are unwilling to open interfaces to customers or other ISPs to fine-tune or improve QoS. Third, not all hardware on current Internet can fully support QoS managements claimed in some protocols like DiffServ [1] and so on. Besides, to support QoS managements, packets have to flow through more procedures, which will decrease the efficiency of these public facilities. End-to-end QoS management [2] is one way to overcome these problems since customers have full controls of endpoints; hence, parameters of these endpoints can be manipulated to support QoS management. In addition, end-to-end QoS managements do not force underlayers to support QoS management; therefore, the efficiency of those public facilities will not be affected. Nowadays, overlay networks are widely used to meet the active demands for multimedia services and end-to-end QoS managements are claimed to be supported on those overlay networks. In this study, we will discuss two aspects of QoS management: QoS qualitative diagnosis and QoS quantitative guarantee, on a peer-to-peer overlay network. QoS diagnosis is usually intended to identify QoS violations [3]. In our approach, we believe causal relationships lie between a QoS metric and its contexts [4] this metric residing in. In this study, we mainly deal with primary contexts, e.g. application configurations, and runtime contexts, like the number of neighbors or buffer map size of a peer in a P2P network. We call these contexts as “soft” contexts with respect to traditional “hard” contexts such as temperature, location and so on. These contexts can be viewed as the causes of the QoS metric since the QoS metric changes along with these contexts. Naturally, we refer QoS qualitative diagnosis to qualitatively identify these causal relationships between a QoS metric and its contexts. In this study, we employ Bayesian network (BN) [5] to exploit these causal relationships as BN is a common approach for representing causal relationships. QoS guarantee tries to satisfy users’ QoS requirements and maintain a QoS metrics on user demanded levels. When the causes of a QoS metric are qualitatively identified by QoS diagnosis, we will show that users’ QoS requirements can be quantitatively fulfilled by tuning these causes or contexts to guarantee the QoS metric. This quantitative guarantee is facilitated by the quantitative representation of BN for causal relationships. Before QoS qualitative diagnosis and QoS quantitative guarantee, QoS metrics and contexts should be discretized first because both BN structure learning and BN parameter learning from continuous variables are time–space consuming tasks. Moreover, redundant relationships may be learned from continuous variables. We will discuss these problems in details in Section 3.1. In this study, we validate our approach on a P2P streaming overlay network scenario through a simulator. We would like to point out that our approach is applicable to more generalized situations or applications because our approach is about a common technique-neutral methodology. Our approach is estimated to have a polynomial time complexity when the number of contexts and QoS metrics involved is held in the order of magnitudes of ten. That is to say our approach is feasible to be put into practice. To sum up, the main contributions of our approach are listed as follows: 1. Firstly, we proposed a systematic approach to qualitatively diagnose QoS and to quantitatively guarantee QoS for the first time. Besides, this approach is a general or common one, which means it is applicable to diagnose and guarantee end-to-end QoS. Our approach can migrate from situations of this study to other applications simply by replacing contexts and QoS metrics of this study with those of target applications. 2. Secondly, we put forward an automatic unwatched discretization algorithm based on Fuzzy set, i.e. Algorithm 2, for discretizing continuous numeric-contexts, which usually follow approximate Gaussian distribution. 3. Thirdly, BN structure learning (K2 algorithm) is employed to exploit causal relationships between contexts and QoS metrics, and a child QoS metric is qualitatively diagnosed to be caused by some parent contexts according to the learned BN structure. An original node-ordering method (see Section 3.2), is proposed to arrange orders for nodes involved in K2 algorithm. 4. Fourthly, BN parameter learning is then used to model those causal relationships quantitatively, and a QoS metric can be guaranteed to a quantitative value by tuning its causal contexts to their corresponding quantitative values according to BN parameter learning. To the best of our knowledge, this is the first work that investigates end-to-end QoS qualitative diagnosis and quantitative guarantee in a systematic way. The remainder of this paper is arranged as follows: a brief summary about related works will be given in Section 2. Our approach will be illustrated in details in Section 3. Simulations and validations will be discussed in Section 4. Section 5 argues the time complexity of our approach. At last, in Section 6 we will make a conclusion about our approach.

The major contribution of our study is that we put forward a systematic approach that brings end-to-end QoS management to current Internet that delivers services with its best efforts. Our approach takes contexts of a service into consideration in a comprehensive manner, which is more flexible than traditional approaches concentrating on one specific metric or factor. Our approach consists of three sequential stages: context value discretization, QoS qualitative diagnosis and QoS quantitative guarantee. We bring forth an automatic unwatched method based on fuzzy logic to discretize continuous contexts. We prove by Theorem 1 that a continuous QoS metric or context usually follows approximate Gaussian distribution. Therefore, our new method uses Gaussian fitting to fit a continuous QoS metric or context, and each term of the fitting result is regarded as the membership function of a discrete value. Our new method requires no prior or expert knowledge; hence, a continuous context can be consistently discretized into the same discrete value whenever and wherever it is discretized. Our experiments show the new method is feasible and effective. Bayesian network theory runs through the whole approach. K2 BN structure learning algorithm is referred to exploit causal relationships between QoS metrics and contexts in QoS diagnosis. A node-ordering algorithm was put forward to arrange the nodes to be processed by k2. The node-ordering algorithm gives orders for nodes in a divide-and-conquer manner. QoS metric and contexts are first classified into three types of sequential groups i.e. primary context group, runtime context group and QoS metric group; then, nodes in each group are sorted into right orders. Bayesian network structure is learned by K2 algorithm with contexts and QoS metric as its nodes. By this BN structure, causal contexts of a QoS metric are parent nodes of the QoS metric node. Simulations on a P2P network simulator show that k2 algorithm has advantages over many other score-based BN structure learning algorithms in both accuracy and efficiency. That is the reason why K2 algorithm is picked up as the BN structure learning algorithm in QoS diagnosis. The optimal number of contexts and QoS metric and the optimal number of samples involved in QoS diagnosis were also discussed w.r.t. accuracy in simulations. Once BN structure is achieved by QoS diagnosis, BN parameter learning is issued to encode causal relationships among BN nodes quantitatively. QoS quantitative guarantee can be made out of a completely defined BN. BN inference is referred to find the largest conditional probability of the QoS metric given its tunable causal (parent) contexts; thus, users’ QoS demands can be quantitatively guaranteed by tuning these causal contexts to their suitable values suggested by BN inference. An example to quantitatively guarantee avg-quality was given in details in simulations. Our approach is estimated to have a polynomial time complexity if static BN inference, the most common situation, is employed. Otherwise, even if dynamic BN inference is referred, our approach still can be finished in a polynomial time complexity if the number of variables, i.e. contexts and QoS metrics, is held in the order of magnitudes of ten. In practice, it is not easy to manage too many variables; in other words, our approach is feasible and deployable. The accuracy of context discretization has a great influence on BN structure/parameter learning. Our automatic unwatched method can help relieve this problem. Our approach works better on synchronized QoS metric and contexts samples. Otherwise, causal relationships exploited by QoS diagnosis may be inaccurate or even incorrect. In addition, the result of BN structure/parameter learning reflects the true characteristics of training data although it may seem strange sometimes. The strange result reveals a strange training data set was used. BN structure/parameter learning or inference was discussed on full observed data in this study. In future, we will extend our approach to deal with incomplete/inconsistent/inaccurate contexts i.e. uncertain contexts since services run on the volatile Internet. Finally, we would like to point out that the scenarios that our approach can work with are far more than P2P network, although P2P related assets, such as P2P network, P2P simulator, P2P service, contexts and QoS metrics of P2P service, are intensively used to illustrate our approach in this study. In our opinion, many topics on P2P QoS research can be considered as a subclass of our context-aware approach since quite a few metrics used in P2P QoS can be thought as contexts just as what we do in this study. Specifically, P2P QoS usually focuses on global optimum and resource efficiency managements on its overlay network; hence, P2P QoS couples with overlay network heavily while our systematic context-aware approach overcomes this problem by taking the metrics of P2P QoS as contexts. Furthermore, our approach works at application layer, which is one of most remarkable features that differ our approach from other approaches that usually work at network layer. From this point, our approach almost depends nothing on the lower layer like network layer. Besides, our approach works in the end-to-end philosophy, that is, it almost can meet with any demands for end-to-end QoS qualitative diagnosis and quantitative guarantee.