Network Service Description and Discovery for ... - Semantic Scholar

4 downloads 140394 Views 217KB Size Report
Feb 8, 2011 - network services to provide networking performance guarantees. .... One of the main challenges to Grid network service description and ...
TAA00002

ACM

(Typeset by SPi, Manila, Philippines)

1 of 17

February 8, 2011

17:1

Network Service Description and Discovery for High-Performance Ubiquitous and Pervasive Grids QIANG DUAN, The Pennsylvania State University Abington College

Ubiquitous and pervasive Grid computing is an emerging computing paradigm that will have a significant impact on the next-generation information infrastructure. Communication networks form a significant integrant of ubiquitous and pervasive Grids and must be utilized effectively by the Grids. The notion of Grid network service may greatly facilitate integrating networking systems into the Grid architecture, and network service description and discovery play a crucial role in the network-Grid integration. Current service description and discovery technologies must be enhanced to meet the special requirements of network service description and discovery for high-performance ubiquitous and pervasive Grids. Network service description needs a model for service provisioning capability and network service discovery must be able to select those networks that meet certain performance requirements. The wide variety of networking systems in ubiquitous and pervasive Grids require general and flexible network service description and discovery approaches that are applicable to heterogeneous networks. The research presented in this article aims at developing network service description and discovery technologies for high-performance ubiquitous and pervasive Grid computing. The main contributions of this article include a general model for describing service capabilities of various networking systems, a service discovery technology for selecting network services that meet the performance requirements specified by Grid applications, and a resource allocation scheme for Grid network services to provide networking performance guarantees. The developed model and technologies are general and flexible; thus are applicable to the wide variety of heterogeneous networks in ubiquitous and pervasive Grid computing environments. Categories and Subject Descriptors: C.2.3 [Computer-Communication Networks]: Network Operations—Network management; C.2.4 [Computer-Communication Networks]: Distributed Systems General Terms: Design, Performance Additional Key Words and Phrases: Network service, service description, service discovery, Quality of Service (QoS), ubiquitous and pervasive Grids ACM Reference Format: Duan, Q. 2011. Network service description and discovery for high-performance ubiquitous and pervasive grids. ACM Trans. Auton. Adapt. Syst. 6, 1, Article 3 (February 2011), 17 pages. DOI = 10.1145/1921641.1921644 http://doi.acm.org/10.1145/1921641.1921644

1. INTRODUCTION

The rapid growth of the Internet, along with the availability of powerful computers and high-speed networks as low-cost commodity components, is changing the way people do computing and manage information. These new technologies have enabled the utilization of a wide variety of distributed computational resources, including computers, storage systems, data centers, and special devices, as a unified resource. This new

Author’s address: Q. Duan, Information Science and Technology Department, The Pennsylvania State University Abington College, 1600 Woodland Rd., Abington, PA 19001; email: [email protected]. Permission to make digital or hard copies part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies show this notice on the first page or initial screen of a display along with the full citation. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component of this work in other works requires prior specific permission and/or a fee. Permission may be requested from Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701, USA, fax +1 (212) 869-0481, or [email protected]. c 2011 ACM 1556-4665/2011/02-ART3 $10.00  DOI 10.1145/1921641.1921644 http://doi.acm.org/10.1145/1921641.1921644

ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

3

TAA00002

3:2

ACM

(Typeset by SPi, Manila, Philippines)

2 of 17

February 8, 2011

17:1

Q. Duan

paradigm that has evolved is popularly termed as “Grid” computing. Over the last decade, significant research efforts and resources have been devoted toward making this vision a reality. Recent advances in computing and communication technologies are rapidly enabling ubiquitous and pervasive Grids where mobile devices and sensing instruments are integrated with classical computing systems. Federation of highly distributed computational resources to deliver better-thanbest-effort services is a key feature of Grid computing, and ubiquitous and pervasive Grids aim at providing computing and communication services anytime and everywhere. Therefore, communication networks with Quality-of-Service (QoS) provisioning capabilities form the basis for ubiquitous and pervasive Grids and should be seamlessly integrated into the Grid architecture. However, until recently there existed a gap between the research in Grid computing and networking fields. The developments of some Grid technologies were not very considerate of the achievable performance of the underlying networks, while research in certain networking technologies was performed with less attention to requirements of Grid computing. Such a gap prevented Grid applications from effectively utilizing the QoS capabilities offered by the underlying network platform. Therefore recently advanced networking for high-performance Grid computing became an important research problem and attracted extensive interest from both industry and academia. Currently the research communities of Grid computing and networking are collaborating on addressing this problem and developing technologies that enable networking resources to be utilized in Grids as easily as other computational resources. The Service-Oriented Architecture (SOA) denotes a set of architectural principles that play a key role in constructing Grid infrastructures. The SOA encapsulates various computing resources in form of services by publishing service descriptions. When a Grid application needs to utilize the resources in a Grid, a service discovery mechanism finds the appropriate Grid service that meets the application requirements. Since networking systems are a type of important resource in Grids, integrating networks into the Grid architecture as a special type of services enables Grid applications to utilize networking resources effectively. Toward this end, the Open Grid Forum (OGF) has formed the “Grid High-Performance Networking” (GHPN) research group and defined the Grid network service as a network service that has roles and/or interfaces that are deemed to be specific to a Grid infrastructure [OGF 2005]. As service description and discovery play a key role in the SOA-based Grid architecture, Grid network service description and discovery form the basis for integrating networking systems into the Grid infrastructure. Grid network services have some special requirements on service description and discovery. The QoS provisioning capability is a key feature that distinguishes different network services. Therefore, the network service description must provide sufficient information about the QoS capability of a networking system and network service discovery should be able to select those networks that can guarantee the QoS performance required by Grid applications. However, the current Grid service description and discovery standards lack effective mechanisms for describing service capability information and selecting services based on their achievable QoS performance. Therefore new approaches for network service capability description and performance-based network service discovery are needed for integrating communication networks into the Grid infrastructure as Grid network services. One of the main challenges to Grid network service description and discovery lies in the heterogeneity of Grid networking systems. In ubiquitous and pervasive Grids a wide variety of networks with heterogeneous networking technologies cooperate with each other dynamically to form the networking platform for various Grid applications. The network service description and discovery mechanisms must be general and ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines)

3 of 17

February 8, 2011

Network Service Description and Discovery for High-Performance Grids

17:1

3:3

independent of network implementations. An end-to-end network connection in ubiquitous and pervasive Grids typically crosses multiple heterogeneous networks, the service description and discovery approaches should be flexible and scalable to support composition of heterogeneous network services. The research presented in this article aims at developing network service description and discovery technologies for high-performance ubiquitous and pervasive Grid computing. The main contributions of this article include a general model for describing service capabilities of various networking systems, a service discovery technology for selecting network services that meet the performance requirements specified by Grid applications, and a resource allocation scheme in Grid network services for network QoS provisioning. The developed model and technologies are general and flexible, thus are applicable to various networks implementations across heterogeneous domains in ubiquitous and pervasive Grid environments. The new approaches for network service description and discovery are complementary to the current Grid service description and discovery standards for improving Grid networking performance. The rest of this article is organized as follows. Section 2 introduces Grid network service for network-Grid integration and discusses the challenges to Grid network service description and discovery. A service capability model for Grid network service description is proposed in Section 3. Section 4 develops the technologies for performance-based discovery of Grid network services and resource allocation in Grid network services for QoS provisioning. Numerical examples are provided in Section 5 to illustrate applications of the developed techniques. Section 6 draws conclusions and discuss possible future works. 2. GRID NETWORK SERVICES FOR NETWORK-GRID INTEGRATION

In the service-oriented Grid architecture, a service is a self-contained implementation of some function(s) with a well-defined interface specifying the message exchange pattern used to interact with the function(s). A Grid service should provide descriptive information about its functions and the required interface for accessing the service. This descriptive information is called a service description, which is published by the service provider typically at a service registry. A Grid infrastructure consists of multiple components that control the utilization of computational resources in the Grid, which are encapsulated in a set of Grid services. One of the most important control components in a Grid is the service broker that discovers Grid services for applications. When an application needs to utilize resources in Grid, the application submits a service request with a requirement specification to a Grid service broker. The service broker searches the service descriptions published at the registry to discover a service that meets the application requirement. Then the service broker retrieves the necessary binding information of the selected service and binds it with the application. The service description and discovery mechanisms for Grid computing are shown in Figure 1. Discovering the appropriate service for meeting each application’s requirement is the key to high-performance Grid computing and service descriptions form the basis for successful service discovery. Therefore, service description and discovery play a crucial role in the service-oriented Grid architecture. The current standards for Grid service description are based on the Web Service Definition Language (WSDL) [W3C 2006]. There is no sufficient descriptive information in WSDL about how well the service can be offered by the provider, for example, the minimum service rate that can be offered to an application. This article refers to such information about a service as the service capability description. The current service discovery in Grid computing is mainly based on Universal Description, Discovery, and Integration (UDDI) specification [OASIS 2005], which lacks an effective mechanism ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines)

4 of 17

February 8, 2011

3:4

17:1

Q. Duan

Fig. 1. Service description and discovery in the service-oriented Grid architecture.

to publish and search nonfunctional features such as service provisioning capabilities. Therefore, the current Grid service description and discovery technologies are mainly function based instead of performance based; that is, the Grid service broker selects a service that supports the required function(s) without considering the achievable QoS performance of the service. For effectively utilizing networking resources in Grids, it is essential to integrate network systems into the Grid architecture as Grid network services. Through Grid network services, networking resources are seen joining other computational resources, such as CPU capacity and storage space, as Grid-managed resources and fully integrated in the Grid architecture. The service-oriented network-Grid integration introduces special requirements on describing and discovering Grid network services. The service capability is a key feature that distinguishes different network services and the network QoS has a significant impact on Grid computing performance. Therefore it is necessary for network service description to provide sufficient information about service capability and for network service discovery to select the Grid network services that guarantee the QoS performance required by Grid applications. Research efforts have been made for enabling service capability description and QoS-based discovery for Grid and Web services. For example, a QoS-capable service broker algorithm was developed to discover, select, and compose Web services that meet end-to-end QoS constraints [Yu and Lin 2004; 2005]. Al-Masri and Mahmoud introduced a relevancy ranking function based on QoS parameters to find the best Web service that meets client QoS preferences [Al-Marsri and Mahmoud 2007]. Technologies for QoS-aware runtime service discovery and selection were proposed in Ambrosi et al. [2005] and Wan et al. [2008], and a QoS model to filter discovered services with their QoS features for maximizing user satisfaction was developed in Li et al. [2008]. Although significant progress has been made in this area, the preceding results were mainly developed for Grid or Web services focusing on data processing and computing instead of data communications and networking, thus may not be applied directly to network service description and discovery. The network research community made significant progress in the past decade toward network QoS provisioning. The integrated services (IntServ) was first developed for offering Internet QoS by reserving resources to traffic flows [Barzilai et al. 1998] and differentiated services (DiffServ) was then developed to enhance scalability of Internet QoS provisioning [IETF 1998]. Many researchers have investigated QoSaware routing algorithms for finding the network routes that meet the performance ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines)

5 of 17

February 8, 2011

Network Service Description and Discovery for High-Performance Grids

17:1

3:5

requirements of multimedia applications. Badia and coauthors [2007] studied QoS routing algorithms in heterogeneous wireless networks. A policy-based QoS support system and a QoS-aware routing algorithm were proposed in Yang et al. [2007]. Various research progress on QoS-routing for supporting QoS in wireless mobile ad hoc networks have also been reported and a survey is given in Chen and Heinzelman [2007]. Recently 3GPP developed the IP Multimedia Subsystem (IMS) [Camarillo and GarciaMartin 2008] and ITU-T developed resources and admission controls functions (RACF) [ITU-T 2008] for controlling QoS in next-generation IP-based networks. The aforesaid research progress focuses on enabling QoS in underlying networking systems and are not considerate of the special requirements of Grid applications. Therefore, there is still a lack of effective interactions between Grid applications and the underlying communication platform to allow network QoS capabilities be fully utilized by Grid applications. The interaction between Grids and networks recently became an active research area. Various research efforts have been made toward monitoring and measuring Grid network characteristics. As examples, Wang and his colleagues developed a SNMP-based monitoring system to measure Grid network characteristics [Wang et al. 2004]. Yousaf and Welzl designed a packet-pair-based measurement method to estimate network path characteristics in a Grid [Yousaf and Welzl 2005]. Derbal defined a Markov-chain-based model for estimating network resource states in Grids [Derbal 2005]. Some of the widely accepted network measurement tools such as Iperf and Network Weather were compared in Yildirim et al. [2008]. These technologies need an effective mechanism that allows the Grid to access and utilize the network state information. The OGF GHPN group is working on a Grid User Network Interface (GUNI) [OGF 2008] and the OGF Network Measurement and Control Working Group (NMC-WG) is developing a protocol for network measurement and control [OGF 2009]. However, how to model network service capabilities for network service discovery and selection are outside the scope of these specifications. Research results on network modeling and description have also been reported. Lacour and his colleagues employed the Directed Acyclic Graph (DAG) to describe network topology and developed a scalable network description model for facilitating Grid application deployment [Lacour et al. 2004]. This description model focuses on a functional view of network topology instead of service provisioning capabilities, thus lacks the information needed for performance-based network service discovery. A Network Description Language (NDL) was developed in van der Ham et al. [2007] as a semantic schema for describing network topology. The NDL language serves more as a vocabulary to present network topology than an approach to model network service capability, and the application of NDL reported in van der Ham et al. [2007] was mainly for optical networks. However, ubiquitous and pervasive Grids typically consist of a wide variety of networks with different implementations. Researchers are also making progress toward discovering network resources and services in ubiquitous and pervasive computing environments. Among these research efforts, different service discovery strategies, including centralized, fully distributed, and semidistributed service discovery, were analyzed and compared in Sivavakeesar et al. [2006]. Zhao and his coauthors [2007] developed a semantic-level service description and applied a fuzzy logic technique to describe imprecise service information. Scalable service discovery frameworks that combine local and remote services in ubiquitous computing systems were developed in Gu et al. [2008] and Chen et al. [2008]. Personalized service discovery technologies were also reported in Frank et al. [2008] and Park and Yoon [2009] to enable discovering services that meet user preferences. Although significant progress has been made in this area, further enhancement is still necessary for the aforementioned results to embrace the ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines)

3:6

6 of 17

February 8, 2011

17:1

Q. Duan

notion of Grid network service for integrating networks into ubiquitous and pervasive Grids. To the author’s best knowledge, little work has been reported with regard to network service capability description and QoS-based discovery mechanisms for Grid network services that are applicable to the heterogeneous networks in ubiquitous and pervasive Grids. The research presented in this article addresses this problem by developing a general model for describing service capabilities of various heterogeneous networks and a technology for discovering network services that meet the QoS performance required by different Grid applications. 3. GRID NETWORK SERVICE DESCRIPTION

A main challenge to network service capability description for ubiquitous and pervasive Grids lies in the heterogeneity of the networking systems in such Grid environments. The heterogeneity in networks requires a general description model that is independent of the architectures and implementation technologies of the described networks. This heterogeneity also implies that end-to-end network connections in ubiquitous and pervasive Grids typically pass through multiple different types of networking systems. Therefore, the network service description model should also be flexible for supporting composition of multiple network services into one composite service. This section proposes a description approach for network service capabilities that meets these requirements. 3.1 Capability Description for Grid Network Services

The main function of a Grid network service is data transportation and the service provisioning capability for transporting data includes two aspects: the destinations that can be reached by the network service and the data transfer capability offered by the network service between each pair of source-destination. In this article, the former aspect is referred to as reachability and the latter aspect is called QoS capability of a network service. Reachability can be described by enumerating all pairs of sources and destinations between which the network service can transfer data. Describing QoS capabilities is more challenging and will be addressed with more detail later in this section. In order to provide a formal and general model for network service capabilities, a Capability Matrix C is defined in this article for describing both reachability and QoS capability for a network service. Given a network service S with m ingress ports and n egress ports, the capability matrix C for this service will be a m × n matrix ⎞ ⎛ c1,1 c1,2 · · · c1,n ⎜ c2,1 c2,2 · · · c2,n ⎟ ⎟ (1) C=⎜ ⎝. . . . . . . . . . . . . . . . . .⎠ , cm,1 cm,2 · · · cm,n where each matrix element ci, j is defined as  0 if the service S cannot transport data from ingress i to ingress j ci, j = Q i, j if the service S provides a network route Ri, j from i to j

(2)

where Q i, j is called the QoS descriptor for the network route Ri, j. According to the definitions given in (1) and (2), the capability matrix element ci, j = 0 if the network service does not support data transportation from the ingress i to the egress j. If the network service provides a network route for transporting data from i to j, then the QoS capability of this route is described by the descriptor Q i, j, which will be further developed in the rest of this subsection. ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines)

7 of 17

February 8, 2011

Network Service Description and Discovery for High-Performance Grids

17:1

3:7

In a Grid network service the routes between different ingress-egress pairs may have different implementations. Even for a particular route, it may pass through multiple heterogeneous networking systems. Therefore, the key requirement for the QoS descriptor is to be independent of network implementations, thus be applicable to various heterogeneous networks in ubiquitous and pervasive Grids. The notion of service curve is adopted from network calculus theory [Boudec and Thiran 2001] to develop a QoS descriptor for data transportation capabilities. The service curve is defined in network calculus as follows. Let B(t) and E(t), respectively, be the accumulated amount of traffic of a flow that arrives at and departs from a server by time t. Given a nonnegative, nondecreasing function, S(·), where S(0) = 0, the server guarantees a service curve S(·) for the flow, if for any time instant t ≥ 0 in the busy period of the server E(t) ≥ B(t) ⊗ S(t),

(3)

where the operator ⊗ denotes

the min-plus convolution operation defined as h(t) ⊗ x(t) = infs:0≤s≤t h(t − s) + x(s) . Essentially a service curve gives the minimum amount of service capacity offered by the server to a client in an arbitrary time interval within a busy period. Therefore a service curve describes the lower bound of the service provisioning capability offered by a service provider to a service customer. A typical server model for networking systems is the Latency-Rate (LR) server [Stiliadis and Varma 1998], which guarantees each flow a service curve βr,θ (t) = r(t − θ ), where θ and r are, respectively, called the latency and service rate parameters for the flow. LR server is a general server model for networks. Traffic controllers widely deployed in practical networking equipments, such as Weighted Fair Queuing (WFQ) and Weighted Round-Robin (WRR), belong to this server category. In the preceding service description approach, the service curve guaranteed by the route Ri, j is adopted as the QoS descriptor Q i, j, which is the element ci, j of the capability matrix C. A service curve is a general descriptive data structure for service capability and is independent of network implementations. Therefore, a service curve is flexible enough to be the capability descriptor for various heterogeneous network routes. In a networking system where a route Ri, j can be modeled by an LR server with a service curve ri, j(t− θi, j), the matrix element ci, j can be represented by a data structure [ri, j, θi, j] with two parameters. Currently there are various mechanisms available for measuring and managing network state information, for example, the technologies reported in Ng and Zhang [2002], Prasad et al. [2003], and Kind et al. [2008], which could be used to obtain network state information for constructing QoS descriptors and building the matrix C. The methods of collecting network state information are network implementation dependent and may vary in different networks, but the matrix C provides all network services a general and standard approach to describe their service provisioning capabilities. 3.2 Capability Description for Network Service Composition

In large-scale ubiquitous and pervasive Grid computing environments, a Grid application typically needs an end-to-end network connection crossing multiple networking systems. If each network is virtualized and described as a Grid network service, the end-to-end data transportation for the application is achieved through a composition of multiple Grid network services, which results in a composite Grid network service. In such a networking scenario the end-to-end network route provided by a composite network service consists of multiple subroutes, each of which is offered by a single Grid network service. Therefore, how to compose the QoS capability descriptors of a set of ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines)

8 of 17

February 8, 2011

3:8

17:1

Q. Duan

heterogeneous subroutes into one descriptor for the end-to-end route is an important and challenging problem. The service curve-based description model developed in the previous subsection supports composition of QoS descriptors. Assume that a service system consists of a series of tandem servers G 1 , G 2 , · · · , G n, which respectively guarantees the service curves S1 (t), S2 (t), · · · , Sn(t) to a flow, it is known from network calculus theory that the service curve guaranteed by the entire system, S(t), can be obtained through the convolution of the service curves guaranteed by each server; that is, S(t) = S1 (t) ⊗ S2 (t) · · · ⊗ Sn(t).

(4)

Since typical networking systems can be modeled as LR servers, we are particularly interested in composition of LR servers. Suppose each network server Si, i = 1, 2, · · · , n, guarantees a service curve βri,θi (t) = ri(t − θi), it can be proved that the convolution of these service curves is n

βr1 ,θ1 (t) ⊗ · · · , ⊗βrn,θn (t) = βr,θ (t),

(5)

where θ = i=1 θi and r = min {r1 , r2 , · · · , rn}. Eq. (5) implies that the if each subroute provided by a single network service guarantees an LR service curve, then the QoS descriptor of the end-to-end route will also be an LR service curve. The total latency of an end-to-end network route equals to the summation of the latency parameters of all its subroutes, and the service rate of the end-to-end route is limited by the subroute with the least service rate. Therefore, suppose a composite Grid network service provides an end-to-end network route R that consists of n subroutes, R1 , R2 , · · · , Rn, whose QoS descriptors are respectively Q 1 , Q 2 , · · · , Q n. Then the QoS descriptor for the end-to-end route R can be obtained from the convolution of the QoS descriptors of all subroutes; that is, Q e = Q 1 ⊗ Q 2 · · · ⊗ Q n.

(6)

Suppose each descriptor Q i is an LR service curve [ri, θi], then the end-to-end route descriptor will be a LR server; that is, Q e = [re , θe ] = [min {r1 , r1 , · · · , rn} ,

n

θi].

(7)

i=1

The capabilities of a network route for QoS provisioning include a variety of state parameters. In addition to the data transportation capacity, security and robustness levels guaranteed by network routes are also important QoS features of Grid network services. The QoS descriptor in the matrix C can serve as a general data structure for describing various QoS metrics. Research in this article mainly focuses on using the QoS descriptor for describing data transportation capabilities of network services and discovering network services that meet bandwidth and delay performance requirements of Grid applications. The QoS descriptor can be extended for including other performance metrics, such as security and reliability levels, into network service descriptions. Such extensions should facilitate composing a set of single service descriptors into one descriptor for a composite service. For example, the security level of a network route provided by a single network service could be described by the probability that the underlying networking security system in the network service is compromised by attacks. For a composite service consisting of multiple network services, the end-to-end network route offered by the composite service is compromised if any one of the subroutes from a single network service is compromised. Since the underlying network infrastructures are operated independently by different network service providers, we can assume that the compromise probability of network routes ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines)

9 of 17

February 8, 2011

Network Service Description and Discovery for High-Performance Grids

17:1

3:9

in different services are independent. Therefore, suppose that an end-to-end network route R consists of n subroutes, R1 , R2 , · · · , Rn, with QoS descriptors Q 1 , Q 2 , · · · , Q n, respectively, and the security parameter for the route Ri is denoted as ki, then the security parameter ke in the end-to-end descriptor Q e can be obtained as ke = 1 −

n

(1 − ki). i=1

Extending the QoS descriptor for including security and reliability features into network service descriptions is an important topic that deserves a systematical study to be presented in a separate paper. Therefore the author limited the scope of the QoS descriptor to data transportation capacities in this article and will explore its extension to other performance metrics in future works. 4. GRID NETWORK SERVICE DISCOVERY

This section will develop a new technology for performance-based discovery of Grid network services which allows the Grid service broker to discover network services that guarantee the QoS performance required by Grid applications. This technology focuses on selecting the appropriate network services while other components of the discovery procedure, including publishing service descriptions, submitting network service requests, and searching the registry for available services, can be implemented based on the current Grid service discovery technologies. 4.1 Networking Demand Profile for Grid Applications

Three aspects of information are needed by a service broker to conduct performancebased network service discovery for a Grid application: (a) the provisioning capabilities of available Grid network services; (b) the networking performance requirement of the Grid application; and (c) the characteristic of network traffic generated by the application. The information (a) can be obtained from the capability matrix C published by the network service provider as part of its service description. The other two aspects of information (b) and (c), which specify the demand of a Grid application on the network service, should be provided to the service broker by the application as part of its service request. Due to the wide variety of applications supported by ubiquitous and pervasive Grids, it is very important to have a general approach for specifying the networking demands ¯ q, ¯ L, ) is defined in this article for of various Grid applications. A Demand Profile P(c, this purpose. This profile consists of three elements: a connectivity set c¯ that specifies the source and destination of data transportation required by a Grid application; a performance requirement set q¯ that consists of the networking performance parameters required by the application; and a traffic load descriptor L. Different parameters may be included in q¯ for different applications, but the minimum bandwidth b req and the maximum delay dreq for data transportation are typical; that is, q¯ = {b req , dreq }. The descriptor L is used to characterize the network traffic that the Grid application will load on a network service. The concept of arrival curve in network calculus theory is employed here for developing a general form of traffic descriptor L for Grid applications. In network calculus, the arrive curve is defined as follows. Let B(t) denote the accumulated amount of traffic generated from an application by time t. Given a nondecreasing, nonnegative function, A(·), the application is said to have an arrival curve A(·) if for all time instants s ∈ (0, t) B(t) − B(s) ≤ A(t − s).

(8)

ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines) 10 of 17

3:10

February 8, 2011

17:1

Q. Duan

Essentially the arrival curve of an application gives an upper bound for the amount of traffic that the application loads on a network service. An arrival curve provides a general description for network traffic load that is independent of any specific application, therefore can be used as the traffic load descriptor L in the demand profile P. Currently most QoS-capable networks apply traffic regulation mechanisms at network boundaries to shape arrival traffic from applications. The traffic regulators most commonly used in practice are leaky buckets. A traffic flow constrained by a leaky bucket has an arrival curve A(t) = min {pt, σ + ρt} , where p, ρ, and σ are, respectively, called the peak rate, the sustained rate, and the maximal burst size of this flow. 4.2 Performance Evaluation for Network Service Selection

The performance evaluation is based on the capability matrix C published by the network service provider and the demand profile P provided by the application. Among various performance requirements, this article focuses on the minimum bandwidth and the maximum delay for data transportation, which are the most significant networking performance parameters for typical high-performance Grid applications. Network calculus provides an effective tool for analyzing QoS performance guaranteed by network services. A service curve itself is a description of the minimum service capacity offered by a network service to a service customer, which essentially gives the minimum bandwidth guaranteed by the network service to an application. Consider a network service S provides a network route R to support a Grid application A. Given that the QoS descriptor for the route R is described by a service curve S(t), then the minimum bandwidth guaranteed by the network service S to the application A on the route R can be determined as bmin = lim [S(t)/t]. t→∞

(9)

Suppose the traffic load descriptor L of the application A is given by an arrival curve A(t), then by following network calculus theory the maximum data transfer delay dmax that the service S guarantees to the application A can be determined from the maximal horizontal distance between the service curve S(t) and the arrival curve A(t); that is,

(10) dmax = max min δ : δ ≥ 0 and A(t) ≤ S(t + δ) . t:t≥0

Since the LR server is a typical network model and the leaky bucket is a widely deployed traffic regulator, performance analysis for a network route modeled by an LR server under traffic load constrained by a leaky bucket regulator is particularly interesting. Assume that the arrival curve for the traffic from a Grid application is A(t) = min{pt, σ + ρt}, and the QoS descriptor of the network route is S(t) = r(t− θ ), then the minimum bandwidth guaranteed to the application on this route is   rθ r(t − θ ) = lim r − = r. (11) bmin = lim t→∞ t→∞ t t Following (10), the maximum data transfer delay guaranteed by the network route to this application can be determined as   p−r σ dmax = θ + (r ≥ ρ). (12) p−ρ r Based on the previously developed performance evaluation technique, the Grid service broker can conduct QoS-based network service discovery through the following process. After receiving a network service request from an application A, the broker ¯ q, ¯ L, ) submitted as part of the request. The broker examines the demand profile P(c, ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines) 11 of 17

February 8, 2011

Network Service Description and Discovery for High-Performance Grids

17:1

3:11

first determines the source and destination for data transfer, which are specified in the connectivity set c¯ of the profile P. Then the service broker will search available network services who have published their service descriptions (including the capability matrix C) at the service registry. Suppose the connectivity set c¯ = {i, j}; that is, transporting data from the source i to the destination j, then a network service is marked as a candidate service only when its capability matrix C has a nonzero element ci, j. In this way, the service broker filters out all network services that do not support data transportation between the source and destination required by the application. Then for each candidate network service S, the service broker employs the previously developed techniques to evaluate the QoS performance that the service S can guarantee to the application A, including the minimum bandwidth b min and the maximum delay dmax . The performance evaluation is based on the QoS descriptor Q i, j provided in the matrix C and the traffic load descriptor L of the profile P. After the performance evaluation, the service broker compares the evaluated results with the application performance requirements, which are given in the set q¯ of the demand profile P, to decide if the service S can be selected for the application A. If the application A has only bandwidth requirement, that is, q¯ = {b req }, then the network service S can be selected for A when b min ≥ b req . If the application A has only delay requirement, that is, q¯ = {dreq }, then the network service S can be selected for A when dmax ≤ dreq . If the application A has both bandwidth and delay performance requirements, that is, q¯ = {b req , dreq }, then the network service S can be selected for A only when b min ≥ b req and dmax ≤ dreq . If there are multiple network services meeting the performance requirements of an application, selection among them may be based on other criteria such as service cost or load balance. The preceding performance evaluation and network selection techniques are applicable to both single and composite network services. For a composite network service, each QoS descriptor in the matrix C is the service curve for an end-to-end route crossing the entire composite service, which can be obtained by following (6) and (7). Therefore, the bandwidth performance evaluated by (11) is the minimum bandwidth that can be offered by the end-to-end network route and the delay performance determined from (12) is the maximum end-to-end delay of the network route through the composite service. Since the service curve-based description of route capacities is independent with network implementations and the arrival curve-based traffic load profile is applicable to various applications, the techniques developed in this subsection apply to heterogeneous networking environments. 4.3 Resource Allocation for Network Service Provisioning

In order to actually guarantee the QoS provisioning to a Grid application, the selected network service must allocate a sufficient amount of resources (mainly bandwidth in the underlying networking systems) for the application. This subsection gives a discussion on bandwidth allocation in Grid network services for QoS provisioning. Eq. (12) shows that given the traffic parameters ( p, ρ, σ ) of an application, the achievable delay upper bound dmax for an application is a decreasing function of the service rate r, which is essentially the amount of available bandwidth that can be allocated on the network route to this application. This implies that the required delay performance can be guaranteed by allocating sufficient amount of bandwidth. Eq. (12) also shows that the minimum possible delay, D min = θ , can be achieved when r = p; that is when the allocated bandwidth is equal to the traffic peak rate. Although allocating bandwidth according to the peak traffic rate achieves the minimum delay, it may cause low utilization in networking resources, especially for applications with fluctuating traffic load. One can also see from (12) that dmax is upper bounded only if r ≥ ρ; ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines) 12 of 17

3:12

February 8, 2011

17:1

Q. Duan

that is, the allocated bandwidth should be at least the sustain rate ρ of the traffic load in order to achieve any delay performance guarantee. Analysis on (12) also shows that dmax is a decreasing function of r that achieves the maximum value D max = θ +σ/ρ when r = ρ. This implies that allocating a bandwidth ra > ρ will guarantee a delay upper bound that is less than D max . Therefore, a reasonable bandwidth allocation scheme is to determine the minimum bandwidth ra between the traffic peak rate p and sustain rate ρ that is sufficient to guarantee the required delay dreq given by the application. That is,

(13) ra = min r : ρ ≤ r ≤ p and dmax (r) ≤ dreq . Given the delay requirement of a Grid application dreq , the network service must guarantee this application a delay upper bound that is no greater than dreq ; that is,   p−r σ dmax = θ + (14) ≤ dreq . p−ρ r Therefore, the minimum amount of bandwidth that must be allocated for satisfying (14) can be obtained as pσ . (15) ra = ( p − ρ)(dreq − θ ) + σ Eq. (15) implies that ra is always less than the peak rate p. As traffic becomes more smooth; that is, when the sustain rate ρ gets closer to the peak rate p, the required bandwidth ra approaches p. In summary, the bandwidth allocation for guaranteeing a delay requirement dreq can be determined as ⎧ ⎪ρ if dreq ≥ D max ⎨ pσ if D min ≤ dreq < D max , ra = ( p−ρ)(dreq −θ)+σ (16) ⎪ ⎩no valid value if d < D req min where D max = θ + σ/ρ and D min = θ . Eq. (16) implies that if the application A requires any loose delay upper bound that is greater than D max , the network service S just needs to allocate an amount of bandwidth that is equal to the traffic sustain rate. For any maximum delay requirement between D max and D min, the bandwidth allocation can be determined from (15). No bandwidth allocation can guarantee any tight delay requirement that is less than D min. If the application has both delay and bandwidth requirements; that is, q¯ = {b req , dreq }, then the minimum amount of bandwidth that must be allocated for this application will be b min = max{ra, b req }. (17) Since the performance analysis developed in the previous subsection applies to both single and composite network services, the preceding resource allocation scheme is applicable to composite network services as well. For composite network services, the bandwidth allocation determined from (16) and (17) is for the end-to-end route. Eq. (7) shows that the actual available bandwidth to a network route crossing multiple network services is limited by the network service that allocates the minimum service rate to the route. Therefore each single service in the composite network service needs to allocate b min to the subroute that it provides to the end-to-end route. 5. NUMERICAL EXAMPLES

Numerical examples are given in this section to illustrate applications of the performance evaluation and bandwidth allocation techniques developed for Grid network ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines) 13 of 17

February 8, 2011

17:1

Network Service Description and Discovery for High-Performance Grids

3:13

Fig. 2. A composite Grid network service that consists of three single network services.

service discovery. Networking scenarios for supporting two Grid applications are considered in the examples. The application A1 uses the network service to deliver a stream of aggregated packet voice and loads the network service with a traffic flow f1 . The application A2 utilizes the network service to transmit a traffic flow f2 for delivering a stream of image data packets. Both A1 and A2 require small maximum packet transmission delay performance. Typical Grid applications utilize both networking and computing services but the examples given in this section mainly focus on the data communications aspect of Grid applications. We assume that the traffic loads generated from both applications are shaped by a leaky bucket regulator when entering the underlying networking infrastructure. The traffic parameters for f1 are peak rate p1 = 3.2 Mb/s, sustained rate ρ1 = 1.12 Mb/s, and the maximum burst size σ1 = 1.13 Mbits. The traffic parameters for f2 are peak rate p2 = 2 Mb/s, sustained rate ρ2 = 87 kb/s, and the maximum burst size σ2 = 1 Mbits. These parameters are derived from the traffic analysis results reported in Butto et al. [1991]. The networking system used in our examples are shown in Figure 2, where a Grid network service is composed of three single network services S1 , S2 , and S3 . The endto-end network route Ri, j consists of three subroutes Ri,k , Rk,l , and Rl, j that are provided by the single network services. We consider the situations in which the network route Ri, j is used for supporting the Grid applications A1 and A2 ; and analyze the delay performance and resource allocation requirements on the same route but for two applications with different traffic characteristics. To reflect the heterogeneity of networking systems in Grid network services, we consider three network services implemented with different networking technologies. The network services S1 and S3 are implemented by wireless access networks and S2 is provided by a fixed backbone network infrastructure. We assume that S1 is a wireless mobile network with a relatively low transmission rate, for example, an IEEE 802.11b wireless Ethernet with 10 Mb/s link capacity; and that the network service S3 is a wireless access network with a medium data rate, such as a WiMAX network with 50 Mb/s link capacity. The maximum transmission units (MTUs) in S1 and S3 are both 1000 bytes. For the backbone network service S2 we assume 1 Gb/s link capacity and 100 bytes MTU. Suppose that the QoS descriptor for each subroute is an LR server. The latency parameter of each subroute mainly comprises the delay factors such as signal propagation delay, packet transmission delay, and packet processing delay at routers/switches. For a network service, suppose the service rate guaranteed by a network route to a traffic flow is r, the link capacity of the route is T, and the MTU in the network service is L, then the maximum delay for transmitting a packet is L/T and the processing delay experienced by the flow on the network route can be estimated as L/r. If the propagation delay is ignorable, then the latency parameter for a network route can be estimated as L/r + L/T. ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines) 14 of 17

February 8, 2011

3:14

17:1

Q. Duan

Fig. 3. The maximum end-to-end delay performance guaranteed to the traffic flow f1 .

Fig. 4. The maximum end-to-end delay performance guaranteed to the traffic flow f2 .

The delay performance guaranteed by the end-to-end network service to the applications A1 and A2 is first analyzed. The maximum end-to-end delay values achieved with various amounts of available bandwidth r on the network route Ri, j are calculated. The results of delay performance for f1 and f2 are plotted in Figure 3 and Figure 4, respectively. These two figures show that the maximum delay for both traffic flows decreases when the available bandwidth increases; that is, the more bandwidth is available on the network route, the tighter is the delay upper bound guaranteed to the application. By comparing these two delay curves we can see that traffic flows f1 and f2 experience different maximum delays with the same amount of available bandwidth. This observation shows that the delay performance guaranteed by a network service to an application is associated with the traffic characteristics of the application as well as the data transportation capacity of the network route. This verifies that it is necessary for a Grid application to submit a traffic load profile to a network service broker in order to discover a network service that meets its delay performance requirement. The amounts of bandwidth that must be allocated in a Grid network service to guarantee a set of maximum delay requirements for the two applications are also calculated. The bandwidth allocation results for f1 and f2 are plotted in Figure 5 and Figure 6, respectively. From these two figures we can see that the amounts of bandwidth allocation for both traffic flows are decreasing functions of the maximum delay requirement. This means more bandwidth must be allocated in the network service to guarantee a tighter delay requirement. Figure 5 and Figure 6 show the amount of bandwidth allocated to a traffic flow for delay performance guarantee is between the traffic peak rate and sustained rate of the flow. Comparing the two curves of ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines) 15 of 17

February 8, 2011

Network Service Description and Discovery for High-Performance Grids

17:1

3:15

Fig. 5. Bandwidth allocation to flow f1 for delay performance guarantee.

Fig. 6. Bandwidth allocation to flow f2 for delay performance guarantee.

bandwidth allocation also shows that applications with different traffic characteristics need different amounts of bandwidth to achieve the same level of delay performance. 6. CONCLUSIONS AND DISCUSSIONS

Communication networks form a significant integrant of ubiquitous and pervasive Grids and must be utilized effectively by the Grids. The notion of Grid network service may greatly facilitate integrating networking systems into the Grid architecture, and network service description and discovery play a crucial role in such network-Grid integration. Current service description and discovery technologies must be enhanced to meet the special requirements of network service description and discovery for highperformance ubiquitous and pervasive Grids. Network service description needs a model for network service provisioning capability and network service discovery must be able to select those networks that can guarantee certain levels of QoS performance. The heterogeneous networking systems coexisting in ubiquitous and pervasive Grids require general and flexible network service description and discovery approaches that are applicable to various network implementations. The research presented in this article aims at developing network service description and discovery technologies for high-performance ubiquitous and pervasive Grid computing. The main contributions of this article include a general model for describing service capabilities of various networking systems, a service discovery technology for selecting network services that meet the performance requirements specified by Grid applications, and a resource allocation scheme for Grid network services to provide network QoS guarantee. The key of the proposed network service model is a ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

3:16

ACM

(Typeset by SPi, Manila, Philippines) 16 of 17

February 8, 2011

17:1

Q. Duan

capability matrix that describes both reachability and QoS capability of a network service. The core of the network service discovery technology includes a general networking demand profile for specifying Grid application requirements and an analysis technique for evaluating the achievable QoS performance of network services. The newly developed approaches for network service description and discovery are complementary to the current Grid service description and discovery standards for improving Grid networking performance. Application of network calculus theory in this article makes the developed model and analysis techniques applicable to the heterogeneous networks in ubiquitous and pervasive Grid computing environments. The QoS descriptor of network services provides a general data structure for describing the capabilities of network routes for QoS provisioning. This article focuses on using the QoS descriptor for describing data transportation capacities to discover network services that meet the bandwidth and delay performance required by Grid applications. Extending the QoS descriptor for including other performance metrics, such as security and robustness levels, is an important topic that will be explored in future works. In this article the QoS descriptor is specified by the deterministic service curve that gives the lower bound of data transportation capacity on a network route. Although such a descriptor enables a service broker to select network services that guarantee the worst-case performance, bandwidth allocation based on such a deterministic bound might cause low resource utilization, especially for applications that can tolerate occasional violations of the worst-case performance requirements. The stochastic service curve model could be applied in future works to improve resource utilization and guarantee statistical QoS provisioning as well. REFERENCES A L -M ARSRI , E. AND M AHMOUD, Q. H. 2007. QoS-Based discovery and ranking of Web services. In Proceedings of the 16th IEEE International Conference on Computer Communications and Networks. A MBROSI , E., B IANCHI , M., G AIBISSO, C., G AMBOSI , G., AND L OMBARDI , F. 2005. A system for predicting the run-time behavior of Web services. In Proceedings of the International Conference on Services Systems and Services Management. B ADIA , L., M IOZZO, M., R OSSI , M., AND Z ORZI , M. 2007. Routing schemes in heterogeneous wireless networks based on access advertisment and backward utilities for qos support. IEEE Comm. Mag. 45, 2, 67–73. B ARZILAI , T. P., K ANDLUR , D. D., M EHRA , A., AND S AHA , D. 1998. Design and implementation of an rsvpbased quality of service architecture for an integrated service internet. IEEE J. Select. Areas Comm. 3, 397–413. B OUDEC, J. L. AND T HIRAN, P. 2001. Network Calculus: A Theory of Deterministic Queueing Systems for the Internet. Springer. B UTTO, M., C AVALLERO, E., AND T ONIETTI , A. 1991. Effectiveness of the leaky bucket policing mechanisms in ATM networks. IEEE J. Select. Areas Comm. 9, 4, 335–342. C AMARILLO, G. AND G ARCIA -M ARTIN, M. A. 2008. The 3G IP Multimedia Subsystem (IMS) 3rd Ed. Wiley. C HEN, L. AND H EINZELMAN, W. B. 2007. A survey of routing protocols that support qos in mobile ad hoc networks. IEEE Netw. Mag. 21, 6, 30–38. C HEN, Y., M AO, W., AND L I , X. 2008. Federation framework for service discovery in ubiquitous computing. In Proceedings of the 11th International Conference on Communication Technology. D ERBAL , Y. M. 2005. Probabilistic resource state estimation in networked environments: The case of computational Grids. In Proceedings of the 3rd Annual Communication Networks and Services Research Conference. F RANK , K., S URACI , V., AND M ITIC, J. 2008. Personalizable service discovery in pervasive systems. In Proceedings of the 4th International Conference on Networking and Services. G U, X., S HI , H., AND Y E , J. 2008. A hierarchical service discovery framework for ubiquitous computing. In Proceedings of the 3rd International Conference on Pervasive Computing and Applications. IETF. 1998. An architecture for differentiated services. ietf rfc 2475.

ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.

TAA00002

ACM

(Typeset by SPi, Manila, Philippines) 17 of 17

February 8, 2011

Network Service Description and Discovery for High-Performance Grids

17:1

3:17

ITU-T. 2008. Resources and admission control functions in the next generation networks. https://www.itu.int/ITU/worksem/ngn/200604/presentation/s3 lu.pdf. K IND, A., D IMITROPOULOS, X., D ENAZIS, S., AND C LAISE , B. 2008. Advanced network monitoring brings life to the wareness plane. IEEE Comm. Mag. 46, 10, 140–146. L ACOUR , S., P EREZ , C., AND P RIOL A, T. 2004. Network topology description model for Grid application deployment. In Proceedings of the 5th IEEE/ACM International Workshop on Grid Computing. L I , M., Y U, B., R ANA , O., AND WANG, Z. 2008. Grid service discovery with rough sets. IEEE Trans. Knowl. Data Engin. 20, 6, 851–862. N G, T. S. E. AND Z HANG, H. 2002. Predicting internet network distance with coordinates-based approaches. In Proceedings of the IEEE Conference on Computer Communications. OASIS. 2005. OASIS specification: Universal description, discovery and integration (UDDI) version 3.0.2. OGF. 2005. Open grid forum GHPN-RG working draft: Grid network services (GNS). May. www.ogf.org. OGF. 2008. Open grid forum GHPN-RG working draft: Grid user network interface (GUNI). February. www.ogf.org. OGF. 2009. Open grid forum NMC-WG working draft: An extensible protocol for network measurement and control. May. www.ogf.org. PARK , K.-L. AND Y OON, U. H. 2009. Personalized service discovery in ubiquitous computing environments. IEEE Pervas. Comput. 8, 1, 58–65. P RASAD, R., M URRAY, M., D OVROLIS, C., AND C LAFFY, K. 2003. Bandwidth estimation: Metrics, measurement techniques, and tools. IEEE Netw. Mag. 17, 6, 27–35. S IVAVAKEESAR , S., G ONZALEZ , O. F., AND PAVLOU, G. 2006. Service discovery strategies in ubiquitous communication environments. IEEE Comm. Mag. 44, 9, 106–113. S TILIADIS, D. AND VARMA , A. 1998. Latency-Rate servers: A general model for analysis of traffic scheduling algorithms. IEEE/ACM Trans. Netw. 6, 5, 611–624. VAN DER H AM , J., G ROSSO, P., VAN DER P OL , R., T OONK , A., AND DE L AAT, C. 2007. Using the network description lanaguage in optical networks. In Proceedings of the 10th IFIP/IEEE International Symposium on Integrated Network Management. W3C. 2006. World wide web consortium (W3C) recommendation: Web service description language (WSDL) version 2. March. http://www.w3.org/TR/wsdl20-primer/. WAN, C., U LLRICH , C., C HEN, L., H UANG, R., L UO, J., AND S HI , A. 2008. On solving QoS-aware service selection problem with service composition. In Proceedings of the 7th Internatinal Conference on Grid and Cooperative Computing. WANG, J., Z HOU, M., AND Z HOU, H. 2004. Providing network monitoring service for Grid computing. In Proceedings of the 10th IEEE International Workshop on Future Trends of Distributed Computing Systems. YANG, K., W U, Y., AND C HEN, H.-H. 2007. Qos-aware routing in emerging heterogeneous wireless networks. IEEE Comm. Mag. 45, 2, 74–80. Y ILDIRIM , E., S USLU, I. H., AND K OSAR , T. 2008. Which network measurement tool is right for you? A multidimensional comparison study. In Proceedings of the 9th IEEE/ACM International Conference on Grid Computing. Y OUSAF, M. AND W ELZL , M. 2005. A reliable network measurement and prediction architecture for Grid scheduling. In Proceedings of the IEEE/IFIP Workshop on Autonomic Grid Networking and Management. Y U, T. AND L IN, K.-J. 2004. The design of QoS broker algorithms for QoS-capable Web services. Int. J. Web Serv. Res. 1, 4, 33–50. Y U, T. AND L IN, K.-J. 2005. Service selection algorithm for Web services with end-to-end QoS constraints. J. Inf. Syst. E-Bus. Manag. 3, 2. Z HAO, D., Y U, Q., AND L I , W. 2007. Service description based on CW for pervasive service discovery. In Proceedings of the 4th International Conference on Fuzzy Logic and Knowledge Discovery. Received June 2009; revised March 2010; accepted July 2010

ACM Transactions on Autonomous and Adaptive Systems, Vol. 6, No. 1, Article 3, Publication date: February 2011.