Chapter 2 Literature Review This chapter presents a literature review on Load balancing based Traffic Engineering, VoIP application, Hybrid Neuro-Fuzzy System, and Intra & Inter Domain Networks. 2.1 Load balancing The traffic distribution algorithm traditionally used for load balancing are Random, Round- Robin (RR), Weight Round-Robin (WRR), Least Load, Least Connection, Fast Response algorithm [8,43]. For the job of assigning loads, Random algorithm distributes traffic to links or servers randomly, regardless of the load on each line. This can leave the system unsteady. The performance of Random algorithm for load balancing is worse than that of RR [8,43]. Round-Robin algorithm is the simplest one, in which traffic will be distributed to links or servers in turn [8,43]. WRR is the same concept of traffic distribution as RR scheme, except that a fractional weight is given to each link depending on the link performance [43]. Least Load algorithm is distributed a new traffic load to one link or server which is the least load [43]. Least Connections algorithm is distributed a new traffic load to one link or server which is the fewest connections [43]. Fast Response algorithm distributes the traffic loads by choosing the one server with the fastest response, which can only be used on servers [43]. There are two major types of the load balancing in IP networks. One is a server load balancing and the other is a router load balancing. 2.1.1 Server Load Balancing RFC 2782, 2000 [29], presented a DNS server load balancing using round-robin method. A user request is distributed in turn of system. James Aweya, 2002 [8], presented an adaptive load balancing scheme for web server which combines admission control and load balancing scheme for distributed web server systems. The admission control mechanism adaptively determines the client request acceptance rate to meet the web servers performance requirements while the load balancing or client request distribution mechanism determines the fraction of requests to be assigned to each web server. Weights based on the server utilization are used to distribute the client requests. RFC 3568, 2003 [9], presented a summary of Request-Routing techniques that are used to direct client requests to surrogates based on various policies and a possible set of metrics. This RFC introduces to use a network latency measurements, Hop Counts, or BGP Information as the metric. Hitoshi Yamada, et al., 2003 [71], proposed a dynamic load balancing to provide both QoS and effectively use network resource in a Content Delivery Network (CDN). The optimum route and server for a user request is selected according to the load of network 4
and content server. The per-flow load balancing distributes traffic loads according to the load of each path which is based on a concept of combination of link and server utilization. In the past time, traditional control algorithms for load balancing of web servers use basically round-robin schemes and have the shortcoming because of the inability to adjust to actual resource usage at the web servers [8]. Thus, the new proposed techniques use basically the approach of weight roundrobin with the dynamic weight. Servers are presented client requests in proportion to their weighting which base on the concept of link utilization, server utilization, or a combination of link and server utilization. 2.1.2 Router Load Balancing In RFC 2328, 1998 [46], Equal-cost multipath (ECMP) distributes equally IP packets among multiple paths. In OSPF-OMP, 1998 [68], suggests using unequal traffic distribution on multiple paths. Nowadays, commercial IP routers have support ECMP and OMP that use typically a per-packet load balancing by round-robin (RR) method [40]. However, the packet-based traffic load balancing would cause a problem of out-ofsequence packet delivery, which will result in the unnecessary retransmission of packets and degradation of throughput in TCP communication [40] Z. Cao, et al., 2000 [13] proposed a flow based IP packet load balancing method. The number of flows is distributed to multiple routes by round-robin method [40]. Since above methods use typically a static link capacity as the link cost for traffic distribution. The traffic allocation is independent of congestion status of each path. Thus it is possible that one of the paths will be more congested than the other [40]. It s suitable to apply RR to a system only when the links have fixed performance levels, since they have no sense of link congestion. K. Leung, et al., 2000 [41] proposed the Call-Based weighted Fair Routing that uses the traffic loads on each link in the network as the concept of the weight for distributing flow traffic. Jun wang, et al., 2003 [69] presented a per-flow load balancing which uses a new link weight function considers both link capacity and hop count for distributing traffic. The weight function is call Inverse-Capacity with Exponential Penalty (ICEP), which a link penalty is related to hop count values. Der-Chiang, et al., 2005 [43] proposed a Dynamic Weighted Round-Robin (DWRR) to determine the values of the weight for traffic distribution in network. The weight is the inverse ratio of link loads. The traffic distribution is the same the concept as the WRR scheme. Mathematical functions are developed to predict each traffic load on the link in the system at time intervals of detection of link loads. This tries to avoid a problem of continue measuring, tracing, ranking, and calculating link loads in the network. Zenghua Zhao, et al., 2005 [7 ] presented an adaptive per-flow load balancing in MPLS network, which uses a path delay as a concept of the weight for distributing traffic. The path delay is estimated by sending the probing packets periodically to each LSP and measuring their round-trip time (RTT). Yoji Kishi, et al., 2005 [40] presented an adaptive per-flow load balancing. Frequency changes in wireless link qualities due to propagation condition and the changes of traffic distribution are used to as a concept of the load balancing. This is specifically designed for wireless environment. In summary, there are two major types of the load balancing network level. One is the packet-level load balancing and the other is the flow-level load balancing. Basically, the proposed load balancing methods distribute traffic loads (packets or flows) to each link depending on the set of weights that bases on the concept of link capacity (bandwidth), link 5
utilization (ratio of link capacity and traffic load on link), frequency changes in wireless link qualities, a combination of link capacity and hop count, or path delay. The main target of load balancing based Traffic Engineering in the network is avoiding link congestion by distrusting traffic loads over multiple paths. Thus, the proposed per-flow load balancing in the network devices is basically based on the concept of the weight round-robin with the dynamic weight because such approach has a sense of link congestion. Up to now, the adaptive load balancing method which is specifically designed for providing VoIP throughput in IP networks is still open. 2.2 Voice over IP (VoIP) RFC 1889, 1996 [58] specifies the real-time transport protocol (RTP), which is an Internet protocol for end-to-end delivery services of real-time data such as audio and video. RTP itself does not guarantee real-time delivery of data; however it does provide mechanisms for the sending and receiving applications to support streaming data [41]. Typically, RTP runs on top of the UDP protocol that provides neither congestion control nor adaptability [65]. ITU-T p.800, 1996 [34] specifies a method called Mean Opinion Score (MOS) for performing controlled evaluation of subjective voice quality using listening tests with untrained subjects. The MOS is a measurement of the quality of human speech at the destination end of the circuit. The outcome of the MOS test is a MOS number that is expressed as a single number in the range of 1 (lowest perceived quality) to 5 (highest perceived quality),. The MOS test is an important factor in determining the QoS for voice communication on the TCP/IP network. Today, the MOS principle has also been applied to gaming on the Internet [64] RFC 2330, 1998 [55] introduces a general metric criteria for the TCP/IP network called IP Performance Metrics (IPPM). RFC 2680, [3] defined a metric for one-way packet loss across Internet paths There are several proposed metrics about packet loss. One is a Type-P-One-way-Packet-Loss that is metric for measuring a single packet loss. The outcome value of such metric is either a zero (signifying successful transmission of the packet) or a one (signifying loss). Another one is a Type-P-One-way-Packet-Loss-Average that is the average of all Type-P-One-way- Packet-Loss values in the Stream. It should be noted that the Type-P-One-way-Packet- Loss-Average is undefined if the sample is empty. RFC 2681, 1999 [4] defines a metric for round-trip-delay of packets across Internet path. There are also several proposed metrics about packet delay. One is a Type-P-Roundtrip-Delay that is a metric for measuring a single packet loss. The result value of a Type-P- Round-trip-Delay is either a real number, or an undefined (informally, infinite) number of seconds. RFC 3393, 2002 [20] defined a metric for variation in delay of packets across Internet paths. The metric is based on the difference in the One-Way-Delay of selected packets. This difference in delay is called "IP Packet Delay Variation (ipdv)". The metric is valid for measurements between two hosts. RFC 3432, 2002 [57] defined a Type-P-One-way-Delay-Periodic-Stream metric that describes a periodic sampling method and relevant metrics for assessing the performance of IP networks. First, the paper motivates periodic sampling and addresses the question of its value as an alternative to Poisson sampling described in RFC 2330. The benefits include applicability to active and passive measurements, simulation of constant bit rate (CBR) traffic (typical of multimedia communication, or nearly CBR, as found with voice activity detection), and several instances where analysis can be simplified. The 6
sampling method avoids predictability by mandating random start times and finite length tests. Following descriptions of the sampling method and sample metric parameters, measurement methods and errors are discussed. Finally, they give additional information on periodic measurements including security considerations. RFC 4656, 2006 [61] defined the One-Way Active Measurement Protocol (OWAMP) which measures unidirectional characteristics such as one-way delay and oneway loss. High-precision measurement of these one-way IP performance metrics became possible with wider availability of good time sources (such as GPS and CDMA). OWAMP enables the interoperability of these measurements. Finally, the IP Performance Metrics Working Group (IPPM WG) provides the set of standard metrics that can be applied to the quality, performance, and reliability of Internet data delivery services. These metrics support a protocol to enable communication among test equipment that implements the one-way metrics. These metrics also can be performed by network operators, end users, or independent testing groups. It should noted that such metrics are not represent a value judgment (i.e. define "good" and "bad"), but rather provide unbiased quantitative measures of performance. ITU-T g.1010, 2001 [38 ] lists the key parameters impacting on user experience as delay, delay variation, and information loss, which are basic requirements for a good quality service of voice communication and ITU-T y.1541, 2002 [37] also includes rate of errored packets to such key parameter [66] RFC 3550, 2003 [59] described RTP which is a real-time application protocol. It is responsible to provide end-to-end network transport functions suitable for real-time application such as audio, video or simulation data, over multicast or unicast network services. This protocol does not address resource reservation and does not guarantee quality-of-service for real-time services. The data transport is augmented by a control protocol (RTCP) to allow monitoring of the data delivery in a manner scalable to large multicast networks, and to provide minimal control and identification functionality. L. Deri, 2005 [21] mentioned to voice over IP (VoIP) that it is much popular as it is a cost effective to reduce telephony costs using the Internet. Although many projects are focusing on developing tools and solutions for building the voice infrastructure, there is very little available in terms of tools and metrics for measuring the impact of VoIP on a network. This paper described the design and implementation of open source tools for detecting and measuring VoIP traffic based on both standard and proprietary protocols. S. Chan, C. Kok, and A. K. wong, 2005 [14] proposed an active buffer management method called Jitter Detection for gateway-based congestion control to stream multimedia traffics in packet-switched network. S. Jelassi and H. Youssef, 2006 [39] presented a playback algorithm denoted PAA (Periodic Adaptive Algorithm) that adapts periodically the playback latency based on the observation of network delay. In summary, problems of VoIP quality are almost always related to network delay, jitter, and packet loss, or some combination of the three. However, there is provides no concrete method to combination of such factors together for VoIP metric. Chapter 4 of this study uses some metric of IPPM (e.g. packet delay calculation) as a basic calculation of the proposed VoIP metric. 2.3 Hybrid Neuro-Fuzzy Traditionally, many authors have defined the methods of neuro-fuzzy integration as two main terms. One is Neuro-Fuzzy and the other is Fuzzy Neural Network. Palade et al., [53] stated that if fuzzy logic is the basic methodology and neural network the subsequent, it is 7
usually termed Neuro-Fuzzy. If the neural network is the basic methodology and fuzzy logic is the second, it is usually termed Fuzzy Neural Network. In [50], Delte suggested that if the fuzzy approach is used to enhance the learning capabilities, or the performance of the neural network, it is termed fuzzy neural network. There are several kinds of possible models for neuro-fuzzy integration, such as using the fuzzy rule to tune the weights of neural network. In this case, fuzzy rules have to be known in advance. Then learning (adjustment weights using the fuzzy rules) takes place in the neural network part alone [49]. Delgado et al. [19] presented an algorithm for training perceptrons using the fuzzy rules. Delte et al. [48] presented a fuzzy perceptron that displays the structure of a multilayer perceptron, however the weights are modeled as a fuzzy set. The intention of this approach is to be able to use prior rule based knowledge, so that the learning does not to have to start from scratch [48]. In chapter 3, the proposed classifier is based on a perceptron whose weights are tuned by a fuzzy system, which enables the classifier to dynamically classify applications according to changes in the current status of application characteristics and resources usage. In the initial state, we did not use the fuzzy perceptron like [48] since this approach is initialized with prior knowledge. On the other hand, the proposed machine needs to train with no a priori knowledge of input-output relations and the neural network method is well studied for this purpose [1]. Thus, we use a learning scheme of neural networks for machine learning in this state. In the adaptive state, the proposed machine uses the fuzzy rules to adjust the weights of neural network like [19] since in this state there are two desired targets. Neural network is basically implemented to one desired target. If there are several targets (conditions), the fuzzy system is well suitable for making decisions from several conditions into a single output [52]. Although, in the initial state the proposed machine uses only the neural network part with learning that needs to start from scratch, it seems to be that the proposed machine should be termed fuzzy neural network since it uses the fuzzy method to improve the performance of a neural network like the definition presented in [50] mentioned above. However, we prefer to refer to the proposed machine as a hybrid neuro-fuzzy to avoid confusing of a structure. Since the proposed machine consists of two machine states, wherein its structure is different from traditional fuzzy neural network, e.g. fuzzy perceptron [48], and the mamdani model [52]. The hybrid neuro-fuzzy is a combination of the neural network and the fuzzy system into one homogeneous architecture [46-47, 53-54]. This system is usually a neural network orientation and can be interpreted as a special neural network with fuzzy parameters, which uses supervised learning [46-47]. In summary, the proposed intelligent load balancing mechanism in the chapter 3 can be classified as the hybrid neruo-fuzzy. 2.4 Intra and Inter Domain IP network An intra-domain IP network is a network under a single administrative domain such as ISPs, where knowledge of the link characteristics and input traffic matrix can be obtained. Typically, Traffic Engineering solutions (e.g., Diffserv) are most effective within the intradomain IP network since all intermediate nodes within the intra-domain IP network can be controlled [30,55-58] as shown in Fig 2.1 8
Fig 2.1 shows the intra-domain IP Network which all intermediated nodes can be controlled. On the other hand, an inter-domain IP network contains two or more intra-domain IP networks. Therefore, there are many different domains (e.g. different ISP networks, enterprise networks, and carrier networks). In the case of inter-domain environments, since each domain is managed according to a specific policy, it is natural that a policy in one domain is different from a policy in another domain. Therefore, it is impossible to seamlessly provide all network information from all intermediate nodes and QoSguaranteed services within the inter-domain IP network e.g. the Internet [28] as shown in fig. 2.2 Fig 2.2 shows the inter-domain IP network which all intermediated nodes can not be controlled In summary, as mentioned in subsection 2.1.2, existing load balancing tools make no attempt to alleviate the serialization delay problem for delay-sensitive application, thus an intelligent load balancing mechanism in the intra-domain IP network is needed. Therefore, chapter 3 of this dissertation presents a possible intelligent load balancing mechanism for the intra-domain IP network. Since all intermediate nodes within the inter-domain IP network can not be controlled, the challenging point in this environment is how to estimate end-to-end VoIP capability paths for each feasible end-to-end path in order to know how each feasible endto-end path should obtain what number of VoIP flows. Thus, chapter 4 of this dissertation also presents a novel VoIP metric as one part of the proposed intelligent load balancing mechanism for the inter-domain IP network. 9