Grid computing


Published on

1 Comment
  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Grid computing

  1. 1. IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 20, NO. 6, JUNE 2008 851 Grid Service Discovery with Rough Sets Maozhen Li, Member, IEEE, Bin Yu, Omer Rana, and Zidong Wang, Senior Member, IEEE Abstract—The computational grid is rapidly evolving into a service-oriented computing infrastructure that facilitates resource sharing and large-scale problem solving over the Internet. Service discovery becomes an issue of vital importance in utilizing grid facilities. This paper presents ROSSE, a Rough sets-based search engine for grid service discovery. Building on the Rough sets theory, ROSSE is novel in its capability to deal with the uncertainty of properties when matching services. In this way, ROSSE can discover the services that are most relevant to a service query from a functional point of view. Since functionally matched services may have distinct nonfunctional properties related to the quality of service (QoS), ROSSE introduces a QoS model to further filter matched services with their QoS values to maximize user satisfaction in service discovery. ROSSE is evaluated from the aspects of accuracy and efficiency in discovery of computing services. Index Terms—Grid computing, Semantic Web, grid service discovery, QoS modeling, Rough sets. Ç1 INTRODUCTIONW ITH the development of Web service technologies [1], the computational grid [2] is rapidly evolving into aservice-oriented computing infrastructure that facilitates and nonfunctional properties (attributes). Advertising services in a grid environment means that service-asso- ciated properties are registered with a service registry.resource sharing and large-scale problem solving over the Service discovery involves a matching process in which theInternet [3]. The Open Grid Services Architecture (OGSA) properties of a service query are matched with that of a[4], promoted by the Open Grid Forum (OGF, http:// service as a standard service-oriented architecture In a grid environment, service publishers may advertise(SOA) for grid applications, has facilitated the evolution. It services independently using their predefined properties tois expected that the Web Service Resource Framework describe services. Therefore, uncertainty of service proper-(WSRF) [5] will be acting as an enabling technology to drive ties exists when matching services. An uncertain property isthis evolution further. The promise of SOA is the enabling defined as a service property that is explicitly used by oneof loose coupling, robustness, scalability, extensibility, and advertised service but does not appear in another serviceinteroperability for large-scale grid systems. advertisement that belongs to the same service category. As shown in Fig. 1, various resources on the Internet This can be further illustrated using Table 1. For example,including processors, disk storage, network links, instru- property P1 , which is explicitly used by service S1 in itsmentation and visualization devices, domain applications, advertisement, does not appear in the advertisement ofand software libraries can be exposed as OGSA/WSRF- service S2 . Similarly, property P3 , which is explicitly usedbased grid services, which are usually registered with a by service S2 , does not appear in the advertisement ofservice registry. A service bus building on service-oriented service S1 . When services S1 and S2 are matched with agrid middleware technologies such as Globus [6] enables service query using properties P1 , P2 , P3 , and P4 , propertythe instantiation of grid services. A grid environment may P1 becomes an uncertain property in matching service S2 ,host a large number of services. Therefore, service dis- and property P3 becomes an uncertain property in match-covery becomes an issue of vital importance in utilizing ing service S1 . Consequently, both S1 and S2 may not begrid facilities. discovered because of the existence of uncertainty of Grid services are implemented as software components, properties even though the two services are relevant tothe interfaces of which are used to describe their functional the query. It is worth noting that properties used in service advertisements may have dependencies, e.g., both P1 and. M. Li is with the School of Engineering and Design, Brunel University, P3 may be dependent properties of P2 when describing Uxbridge, UB8 3PH, UK. E-mail: services S1 and S2 , respectively. Both S1 and S2 can be. B. Yu is with Level E Limited, ETTC, The King’s Buildings, Mayfield Road, Edinburgh, EH9 3JL, UK. E-mail: discovered if P1 and P3 (which are uncertain properties in. O. Rana is with the School of Computer Science, Cardiff University, terms of the user query) are dynamically identified and Queen’s Buildings, 5 The Parade, Roath, Cardiff, CF24 3 AA, UK. reduced in the matching process. To increase the accuracy E-mail: Z. Wang is with the School of Information Systems, Computing, and of service discovery, a search engine should be able to deal Mathematics, Brunel University, Uxbridge, UB8 3PH, UK. with uncertainty of properties when matching services. E-mail: In this paper, we present ROSSE [21], [22], [23], [24]: aManuscript received 10 Apr. 2007; revised 20 Sept. 2007; accepted 6 Dec. search engine for grid service discovery. Building on Rough2007; published online 19 Dec. 2007. sets theory [25], ROSSE is novel in its capability to deal withFor information on obtaining reprints of this article, please send e-mail to: uncertainty of service properties when matching, and reference IEEECS Log NumberTKDE-2007-04-0151. This is achieved by dynamically identifying and reducingDigital Object Identifier no. 10.1109/TKDE.2007.190744. dependent properties that may be uncertain properties 1041-4347/08/$25.00 ß 2008 IEEE Published by the IEEE Computer Society
  2. 2. 852 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 20, NO. 6, JUNE 2008 TABLE 1 Two Service Advertisements with Uncertain Service Properties inference engine to infer the semantic relationships of properties when matching services. Service discovery. A user posts a service query to ROSSE via its Web user interface (step 5). The query includes a service category of interest and expected service properties. The query is then passed to the Irrelevant Property Identification component (step 6), which accesses the ROSSE Service Repository (step 7) to identify and mark theFig. 1. A layered structure of service-oriented grid systems. properties of advertised services that are irrelevant to the properties used in the service query based on the ontologieswhen matching a service query. In this way, ROSSE defined in the ROSSE Ontology Repository (step 8). Theincreases the accuracy in service discovery. In addition, query is then passed to the DPR component (step 9), whichfunctionally matched services may have distinct nonfunc- accesses the ROSSE Service Repository to identify and marktional properties related to the quality of service (QoS). To dependent properties (step 10). Upon completion, the DPRmaximize user satisfaction in service discovery, ROSSE component invokes the Service Similarity Computing (SSC)introduces a QoS model to further filter matched services component (step 11), which accesses ROSSE Service Repo-with their QoS values. Finally, ROSSE is evaluated from the sitory (step 12) to compute the match degrees of relevantaspects of accuracy and efficiency in discovery of comput- properties of advertised services to the service query. Aning services. irrelevant property is given a match degree of zero. The SSC The remainder of the paper is organized as follows: component further computes the similarity degrees ofSection 2 presents the design of ROSSE with a focus on advertised services to the service query using the matchdependent property reduction (DPR). Section 3 introduces a degrees of their individual properties. It should be notedQoS model to filter matched services with their QoS values. that dependent properties that may be uncertain propertiesSection 4 briefly describes the implementation of ROSSE are not involved in the similarity computing process. As aand gives a case study to illustrate the application of ROSSE result, the similarity degrees of advertised services will notfor discovery of computing services. Section 5 evaluates the be affected by these uncertain properties. In this way, ROSSEaccuracy and efficiency of ROSSE in service discovery. can discover the services that are most relevant to the serviceSection 6 discusses some related work, and Section 7 query. Up to now, advertised services are matched withconcludes the paper. their functional properties. As functionally matched services may have distinct nonfunctional properties related to QoS, the SSC component invokes the QoS Modeling component2 THE DESIGN OF ROSSE (step 13), which in turn filters functionally matched servicesROSSE considers input and output properties individuallywhen matching services. For the simplicity of expression,input and output properties used in a service query aregenerally referred to as service properties. The same goesfor service advertisements. Fig. 2 shows ROSSE compo-nents. The interactions between the components follow twoprocesses—service publication and service discovery. Service publication. Service publishers advertise theirservices to ROSSE through a Web user interface (step 1).Advertised services with WSDL interfaces or OWL-S [13]interfaces are then loaded into the ROSSE Service Reposi-tory, in which the elements of services such as the namesand properties of services are registered with ROSSE(step 2). When advertising services, service publishersmay also publish service ontologies that can be defined inOWL [14]. These OWL ontologies are then parsed by anOWL parser (step 3) and loaded into the ROSSE OntologyRepository (step 4). The ontology repository is used by an Fig. 2. ROSSE components.
  3. 3. LI ET AL.: GRID SERVICE DISCOVERY WITH ROUGH SETS 853with QoS values (step 14). Finally, a list of discovered We define the following relationships between pQ and pAservices that are ranked with their functionally matched based on the work proposed by Paolucci et al. [15]:degrees is presented to a user (step 16) via the Web userinterface of ROSSE (step 15). Each of the discovered services . Exact match. pQ and pA are equivalent, or pQ is ahas a QoS value associated with it. subclass of pA . In the following sections, we describe in depth the . Plug-in match. pA subsumes pQ . . Subsume match. pQ subsumes pA .processes involved in service discovery in ROSSE. First, we . Nomatch. There is no subsumption between pQintroduce Rough sets for service discovery. and pA .2.1 Rough Sets for Service Discovery If pA has a nomatch relationship with each pQ used in aThe Rough sets theory can be considered as a mathematical service query, then pA will be marked as an irrelevanttechnique to deal with uncertainty in knowledge discovery property when matching the service query.[35]. A fundamental principle of a Rough sets-based As introduced in Section 1, uncertainty of properties maylearning system is to discover redundancies and depen- exist in advertised services when matching a service query.dencies between the given features of a problem to be An uncertain property is a service property that is explicitlyclassified. The Rough sets theory approaches a given used in one advertised service but does not appear inconcept using lower and upper approximations. another service advertisement that belongs to the same Let service category. As advertised services are structured as service records in ROSSE using a database, we define an . be a domain ontology, uncertain relationship between pQ and pA as follows: . U be a set of N advertised services whose properties are defined in , U ¼ fs1 ; s2 ; . . . ; sN g, N ! 1, . Uncertain. There is no subsumption between pQ and . P be a set of K properties that describe the N pA , and pA has an empty value, which is NULL. advertised services of the set U, P ¼ fp1 ; p2 ; . . . ; pK g, K ! 1, 2.3 Dependent Property Reduction . PA be a set of M properties that are relevant to the The properties of advertised services may have dependen- properties used in a service query Q in terms of , cies by which ROSSE deals with uncertainty of properties. PA ¼ fpA1 ; pA2 ; . . . ; pAM g, PA P , M ! 1, Dependent properties are indecisive (redundant) properties . X be a set of advertised services that are relevant to that can be reduced when matching services. A reduct is a the service query Q in terms of , X U, set of decisive properties that are sufficient enough to . X be a lower approximation of the set X, describe those advertised services that are relevant to a . X be an upper approximation of the set X, and service query. Based on the Rough sets theory, ROSSE . ½xŠPA be a set of advertised services that are identifies indecisive properties in the following way: exclusively defined by the properties of the set PA , Let x 2 U. . , X, P , and PA be defined as in Section 2.1, According to the Rough sets theory, we have D . PA be a set of LD decisive properties when X ¼ fx 2 U : ½xŠPA Xg; ð1Þ matching a service query Q in terms of , PA ¼ fpD ; pD ; . . . ; pD D g, PA PA , LD ! 1, D A1 A2 AL D IND X ¼ fx 2 U : ½xŠPA X 6¼ ;g: ð2Þ . PA be a set of LIND indecisive (dependent) properties when matching the service query Q in For a service property p 2 PA , we have the following: terms of , . 8x 2 X, x definitely has property p. PA ¼ fpIND ; pIND ; . . . ; pIND g; IND A1 A2 ALIND . 8x 2 X, x possibly has property p. IND PA PA ; LIND ! 1; . 8x 2 U À X, x absolutely does not have property p. For a service query, there could be a large number ofmatched services. Using the size of the set X, a user can . INDðÞ be an indiscernibility relation,dynamically determine the size of the set X that would . f be a mapping function from a property to an advertised service,maximize user satisfaction in service discovery. The . SðQ; sÞ be the similarity degree of an advertisedselection of services based on lower and upper approxima- service s to the service query Q in terms of , s 2 Xtions will be further discussed in Section 2.5. (SðQ; sÞ can be computed using (6) to be presented in2.2 Irrelevant Property Identification Section 2.4),The properties used in a service advertisement may have . Y be a set of objective services that are relevant to thesemantic relationships with the properties used in a service service query Q, Y X (we define Y ¼ f8ðx; yÞ 2query based on the definition of a domain ontology. Y ; SðQ; xÞ ¼ SðQ; yÞ : 8s 2 ðX À Y Þ; SðQ; sÞ SðQ; xÞg), Let . ½Y ŠPA be a set of objective services that are defined by the properties of the set PA , . pQ be a property used in a service query and . ½Y ŠP D be a set of objective services that are defined by A . pA be a property used in a service advertisement. D the decisive properties of the set PA , and
  4. 4. 854 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 20, NO. 6, JUNE 2008 . ½Y ŠðPA ÀP IND Þ be a set of objective services that are approach is that the mapping from a semantic relationship A IND defined by the properties of the set PA À PA . to a fuzzy variable does not consider the semantic distances Then, we have of the properties involved. For example, two plug-in IND matches with different semantic distances are mapped to INDðPA Þ ¼ fðx; yÞ 2 X : the same fuzzy variable that is relevant. To increase the ð3Þ 8pIND 2 PA ; fðx; pIND Þ ¼ fðy; pIND Þg; Ai IND Ai Ai accuracy in assigning matching degrees between pQ and pA , semantic distances should be taken into account. ½Y ŠPA ¼ ½Y ŠP D ¼ ½Y ŠðPA ÀP IND Þ : ð4Þ Let A A Expression (3) shows that indecisive properties are . domðpQ ; pA Þ be the degree of a match between pQdispensable in differentiating advertised services, and (4) and pA andfurther indicates that the set of objective services can always . kPQ ; PA k be the semantic distance between pQ and pAbe identified regardless of the existence of indecisive in terms of a domain ontology .properties. Algorithm 1 shows the discovery of decisive Following the work proposed in [36] for assigning matchproperties in which lines 1-7 are used to identify individual degrees, we further define domðpQ ; pA Þ as follows:indecisive properties, and lines 8-15 are used to check all 8possible combinations of these individual indecisive prop- 1 1 exact match; 1 2 þ ðkPQ ;PA kÀ1Þ plugin match; kPQ ; PA k ! 2;erties with an aim to compute a maximal set of indecisive e domðPQ ; PA Þ ¼ 1properties. It should be noted that some uncertain proper- subsume match; kPQ ; PAk ! 1; 2ÂeðkPQ ;PA kÀ1Þ 0:5ties of advertised services may be indecisive properties. As uncertain match; :a result, these uncertain properties will be reduced when 0 nomatch:computing the similarity degrees of services. In other ð5Þwords, for a service query, some uncertain properties may According to (5), for a plug-in match between pQ and pA ,not affect the similarity degree of an advertised service. domðpQ ; pA Þ 2 ð0:5; 1Þ. For a subsume match between pQ andAccordingly, ROSSE increases the accuracy of service pA , domðpQ ; pA Þ 2 ð0; 0:5Š.discovery. Let D . PA and SðQ; sÞ be defined as in Section 2.3, . PQ be a set of M properties used in a service query Q, PQ ¼ fpQ1 ; pQ2 ; . . . ; pQM g, M ! 1, and . domðpQi ; pAj Þ be a match degree between PQi and PAj in terms of a domain ontology , pQi 2 PQ , 1 i M, D pAj 2 PA , 1 j LD . As every decisive property of an advertised service s has a maximal match degree when matching all the properties used in a service query, SðQ; sÞ can be computed using the following: , XX LD M SðQ; sÞ ¼ maxðdomðpQi ; pAj ÞÞ LD : ð6Þ j¼1 i¼1 Therefore, each advertised service has a similarity degree to a service query.2.4 Computing Similarity Degrees 2.5 Lower and Upper Approximations of MatchedAs described in Section 2.2, a match between pQ and pA can Servicesbe exact, plug-in, subsume, uncertain, or nomatch. For each For a service query, the number of matched services couldmatch, a numerical degree should be assigned so that the be large. To facilitate users in choosing the services thatrelationship between pQ and pA can be quantified. Tsetsos would maximally satisfy their queries, the set of discoveredet al. [37] use fuzzy set theory to evaluate service services need to be dynamically determined. We apply thematchmaking. The relationships between the properties of concept of lower and upper approximations of Rough setsadvertised services and the properties of service queries are for this purpose.mapped to fuzzy linguistic variables, e.g., exact is mapped Letto very relevant, and plug-in is mapped to relevant. In this . U, X, X, and X be defined as in Section 2.1,work, linear trapezoidal membership functions are as- . SðQ; sÞ be defined as in Section 2.3,sumed for capturing the vagueness of the various linguistic . jXj be the cardinality of the set X,terms. The preliminary results show the effectiveness of the . jXj be the cardinality of the set X, jXjfuzzy linguistic approach in quantifying match degrees for . be an approximation degree, ¼ jXj , 0 1,service matchmaking. However, one major concern with the jXj 6¼ 0,
  5. 5. LI ET AL.: GRID SERVICE DISCOVERY WITH ROUGH SETS 855 . X1 be a set of advertised services, 8s 2 X1 , SðQ; sÞ ¼ description, and Ninvoke ðsi Þ is the total number of 100 percent, invocations of service si , . jX1 j be the cardinality of the set X1 , . Emax be the maximal successful execution rate . X2 be a set of advertised services, 8s 2 X2 , 0 among the list of SðQ; sÞ 100 percent, and . jX2 j be the cardinality of the set X2 . fE1 ; E2 ; . . . ; En g; Emax ¼ maxðE1 ; E2 ; . . . ; En Þ; Then, according to (1) and (2), we have and . Emin be the minimal successful execution rate among X1 jX1 j 0; the list of X¼ ð7Þ X X2 : jXj ¼ jX2 j  jX1 j ¼ 0; fE1 ; E2 ; . . . ; En g; Emin ¼ minðE1 ; E2 ; . . . ; En Þ: X1 jX1 j 0; X¼ ð8Þ Then, we define X2 jX1 j ¼ 0: Ei ÀEmin If the set X1 exists, then the set X1 will be presented to a Emax 6¼ Emin ; qreliability ðsi Þ ¼ Emax ÀEmin ð9Þuser as both the upper and the lower approximation sets of 1 Emax ¼ Emin :matched services for the service query. It should be noted that running a service reliably If the set X1 does not exist, then a user can apply an demands a certain amount of CPU processing power andapproximation degree to dynamically determine the memory space. The reliability of a service is largely affectedlower approximation set of matched services. In this way, by the capacity of the grid that hosts the service. A grid is aadvertised services with low similarity degrees may not bepresented to the user. The set X2 will be used as the upper dynamic computing environment in nature. For example,approximation set of matched services for the service query. computing nodes may join or leave the environment dynamically. Services can be dynamically deployed to a3 QOS MODELING certain node in the environment upon request. The work- load of a node may change frequently. Considering theAs described in Section 2, ROSSE uses the functional dynamic nature of grid environments, we further tune theproperties of services to match services. However, servicesmay have distinct nonfunctional properties related to QoS. computed qreliability ðsi Þ of service si in such a way thatTo maximize user satisfaction in a service query, function- qreliability ðsi Þ ¼ K  qreliability ðsi Þ. K is a variable that can beally matched services should be further filtered with their computed using K ¼ nnode ðsi Þ , where nnode ðsi Þ is the number NnodeQoS properties. Zeng et al. [26] propose a set of QoS of computing nodes in a grid environment that meet bothproperties for Web service composition. In this section, we the CPU and memory requirements of service si at the timerevisit these QoS properties in terms of grid computing when the service is requested, and Nnode is the total numberenvironments. We classify QoS properties into two clas- of computing nodes in the grid at that time. The resourceses—system-related properties and non-system-related information on the usage of CPU and memory of comput-properties. ing nodes in a grid can be collected using a monitor system3.1 System-Related QoS Properties such as Ganglia.1The performance of a service is largely affected by the 3.1.2 Execution Efficiencycapacity of a computing environment that hosts the service.We discuss three QoS properties in this category, i.e., The execution efficiency qefficiency ðsÞ of a service s refers to how fast the service can be executed. The qefficiency ðsÞ ofreliability, execution efficiency, and availability. service s is measured in terms of its execution duration,3.1.1 Reliability which could be computed using the approach proposed in [26]. The execution duration of a service is the sum of theThe reliability qreliability ðsÞ of a service s represents the ability processing time and the transmission time. It is worthof the service to perform its required functions under stated noting that computing the processing time of a service in aconditions for a specified period of time [40]. In ROSSE, the dynamic grid environment is a challenging issue itself [41],qreliability ðsÞ of service s is measured with its successful [42]. The work presented in [26] does not provide anexecution rate, which can be computed using the statistical approach to compute the processing time of a service. Inapproach proposed in [26]. ROSSE, we measure the execution duration of a service and Let compute its execution efficiency in the following way: Let . S be a set of services that are functionally matched to a service query, S ¼ fs1 ; s2 ; . . . ; sn g, n ! 1, . S be a set of services that are functionally matched to . Ei be the successful execution rate of service si ðsi 2 a service query, S ¼ fs1 ; s2 ; . . . ; sn g, n ! 1, SÞ for a given period of time, Ei ¼ Nnexe ðsðsÞi Þ , where invoke i . Ti be a measured execution duration of service nexe ðsi Þ is the number of times that service si has si ðsi 2 SÞ running on a dedicated computing node been successfully completed within the maximum expected time frame as specified in the service 1.
  6. 6. 856 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 20, NO. 6, JUNE 2008 (each service runs exclusively on the node during its QðsÞ ¼ w1 qavailability þ w2 qreliability execution), þ w3 qefficiency þ w4 qcos t effectiveness þ w5 qreputation ; . Tmax be the maximal execution duration among the ð12Þ list of fT1 ; T2 ; . . . ; Tn g, Tmax ¼ maxðT1 ; T2 ; . . . ; Tn Þ, and where w1 , w2 , w3 , w4 , and w5 are the weights of . Tmin be the minimal execution duration among the availability, reliability, execution efficiency, cost-effective- list of fT1 ; T2 ; . . . ; Tn g, Tmin ¼ minðT1 ; T2 ; . . . ; Tn Þ. ness, and reputation of service s, respectively, wi 2 ½0; 1Š, P5 Then, we define i¼1 wi ¼ 1. The weights are assigned based on user T ÀT preferences. max i Tmax 6¼ Tmin ; qefficiency ðsi Þ ¼ Tmax ÀTmin ð10Þ 1 Tmax ¼ Tmin : 4 ROSSE CASE STUDY3.1.3 Availability In this section, we briefly describe the implementation ofThe availability qavailability ðsÞ of a service s is defined as the ROSSE. Then, we give a case study to illustrate theprobability that service s can be accessed at a particular application of ROSSE to discover computing services.time [26]. In ROSSE, qavailability ðsÞ is computed using 4.1 ROSSE Implementationqavailability ðsÞ ¼ NinvokeðsÞ . For a given period of time, navail ðsÞ navail ðsÞ ROSSE is implemented as a Web system using Java andis the number of successful invocations of service s within Web technologies. ROSSE has a Web user interface, asthe maximum expected time frame as specified in service shown in Fig. 3, for service publication and discovery.description, and Ninvoke ðsÞ is the total number of invocations ROSSE provides users with graphical user interfaces forof service s. publishing services with WSDL interfaces. ROSSE uses the OWL-S API to directly register services with OWL-S3.2 Non-System-Related QoS Properties interfaces. The ROSSE Service Repository consists of aWe classify non-system-related QoS properties into cost- UDDI registry for WSDL services and a service repositoryeffectiveness and reputation. for OWL-S services. jUDDI and mySQL are used to build the UDDI registry. The OWL-S service repository is also3.2.1 Cost-Effectiveness structured as a database that records service elements suchThe cost-effectiveness qcost effectiveness ðsÞ of a service s refers as service names and service the effectiveness of cost in using the service. In ROSSE, To facilitate the reduction of dependent properties,cost-effectiveness is computed in the following way: service records are structured in such a way that each Let column has only one property associated with it. Ontologies defined in OWL documents are loaded into ROSSE using . S be a set of services that are functionally matched to ´ ´ the Protege OWL API, which in turn invokes RACER [27] to a service query, S ¼ fs1 ; s2 ; . . . ; sn g, n ! 1, infer a semantic relationship between two properties . Ci be the cost of using service si , si 2 S, 1 i n, defined in an OWL ontology. However, we implemented . Cmax be the maximal cost among the list of a light-weighted reasoner in ROSSE to replace RACER. The fC1 ; C2 ; . . . ; Cn g, Cmax ¼ maxðC1 ; C2 ; . . . ; Cn Þ, and main reason for this is that RACER has a high overhead . Cmin be the minimal cost among the list of when parsing multiple OWL documents. Each time RACER fC1 ; C2 ; . . . ; Cn g, Cmin ¼ minðC1 ; C2 ; . . . ; Cn Þ. parses an OWL document, it needs an initialization process Then, we define that is time consuming. It should be pointed out that ROSSE C ÀC services that belong to the same service category may be max i Cmax 6¼ Cmin ; defined with distinct ontologies. For example, two ontolo- qcos t effectiveness ðsi Þ ¼ Cmax ÀCmin ð11Þ 1 Cmax ¼ Cmin : gies used by ROSSE may have the same set of service properties but with different topologies. ROSSE manages3.2.2 Reputation multiple ontologies in its ontology repository. WhenThe reputation qreputation ðsÞ of a service s is a ranking degree requested, ROSSE only loads the ontology repository once,of a user’s experience in using the service. In ROSSE, and it maintains multiple ontologies in memory. Theqreputation ðsÞ is computed using the following: relationships of the properties defined in these OWL ontologies are kept in a database. The ROSSE reasoner Pk j¼1 RDj ðsÞ accesses these database records to make inferences between qreputation ðsÞ ¼ ; multiple ontologies. We compare the overhead of RACER k with that of the ROSSE reasoner in Section 5.where RDj ðsÞ is a ranking degree of user j on using services, 0 RDj ðsÞ 1, and k is the total number of users 4.2 Discovery of Computing Services in ROSSEinvolved in the evaluation process. Fig. 4 shows the ontologies used in this case study defining the classifications of grid entities, security, virtual organiza-3.3 Overall QoS Values of Functionally Matched tion (VO) management, computing resource properties, Services CPU properties, and hard-disk properties, respectively.Based on the computed values of the aforementioned QoS When matching properties related to VO management,properties, the overall QoS value QðsÞ of service s can be two ontologies (strictly two ontology topologies) are used tocomputed using the following: classify VOs, which are represented respectively by c1-c4
  7. 7. LI ET AL.: GRID SERVICE DISCOVERY WITH ROUGH SETS 857Fig. 3. The Web user interface of ROSSE.and g1-g3. Computing services are registered with theROSSE Service Repository. In the following parts, wedescribe how services are matched and discovered usingthe following query:4.2.1 Building a Decision TableA service decision table is used to identify dependent Fig. 4. Ontologies used in searching for computing among services. As the number of servicesregistered with ROSSE is large, the decision table isconstructed by sampling registered services. For a specific 4.2.3 Computing Similarity Degreesquery, ROSSE randomly selects a certain number of service Decisive properties are used for computing the similarityrecords. An advertised service record is selected if one of its degrees of advertised services to a service query. A matchproperties has a valid relationship with a property used in a degree can be computed between each decisive propertyservice query. The relationship can be exact, plug-in, or used in a service advertisement and a property used in thesubsume, as described in Section 2.2. service query using the ontologies defined in Fig. 4 and (5), Table 2a shows a segment of the decision table with13 advertised service records. As can be seen in Table 2a,the properties of advertised services that are relevant to the TABLE 2 (a) A Segment of the Decision Table; (b) Identified Dependentservice query are d3, b4, e4, f3, f5, d8, d6, f2, c4/g3, e1, and b3. Properties; (c) the Decision Table without DependentIf a property in a service record is marked with 1, this Properties; (d) Match Degrees to the Service Querymeans that the property is explicitly used by the service inits advertisement. For example, the service S1 has propertiesof d3, b4, e4, f3, e1, and b3 in its advertisement. A propertymarked with X in a service record means that the servicedoes not explicitly have the corresponding property in itsadvertisement. It should be noted that a property markedwith X in a service record does not necessarily mean thatthis property is not relevant to the service. Such a propertycould be dependent on other properties used by the service.ROSSE considers properties marked with X as uncertainproperties when matching services.4.2.2 Reducing Dependent PropertiesOnce a service decision table is constructed, the next step isto identify dependent properties. Using Algorithm 1,presented in Section 2.3, we identify that properties b4, e4,and f3 are dependent (indecisive) properties that can bereduced from the decision table when computing thesimilarity degrees of services. Table 2b shows the depen-dent properties identified, and Table 2c shows the segmentof the decision table without dependent properties.
  8. 8. 858 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 20, NO. 6, JUNE 2008 TABLE 3 lowest similarity degree because it only supports an exact Some Results of ROSSE in Matching Services match. ROSSE performs best because of its reduction of dependent properties (i.e., P1 and P4). OWL-S matching cannot deal with uncertain properties. As a result, OWL-S matching produces a similarity degree of zero when matching services S2-S5 with uncertain properties. In the case of service S5, UDDI produces a match degree of zero because there is no exact match in the service advertisement. ROSSE shows its effectiveness in dealing with uncertain properties when matching services S2-S5 with a match degree of 96 percent, 79 percent, 91 percent, and 62 percent, respectively. 5.1.2 Measuring Precision and Recallpresented in Section 2.4. Table 2d shows match degrees of Precision and recall are standard measures that have beenthe decisive properties used in the 13 service records. It used in information retrieval for measuring the accuracy ofshould be noted that both c1 and g1 refers to the same a search method or a search engine [38], [39]. To evaluateproperty vomanage, but they are defined with different the precision and recall of ROSSE in service discovery, weontology topologies. The match degree of the vomanage selected a set of 30 services, of which 10 services wereproperty used in the query to the sharedresource property relevant to a service query. Each service had five properties,used in advertised services is computed in such a way that a of which two were dependent properties. We performedmean of two match degrees using the two ontology two groups of tests, of which each group involved 10 tests.definitions (i.e., 50 percent and 18 percent) is computed, For each test, we produced a list in which the 30 serviceswhich is 34 percent. were listed in a random order. In the tests of group 1, we It is worth noting that for an uncertain property that is enforced two constraints on the selected services. First, nomarked with X in Table 2d, a match degree of 50 percent service had an uncertain property. Second, at least oneis given using (5), presented in Section 2.4. The similarity property of a service was assigned an exact match. Thisdegree of an advertised service to a service query can be ensured that all the relevant services were returned bycomputed using (6), presented in Section 2.4. For the UDDI matching, OWL-S matching, and ROSSE, respec-service query, for example, service S1 has a similarity tively. In the tests of group 2, we removed the twodegree of 67 percent, and service S13 has a similarity constraints. We allowed a few services in each of the 10 listsdegree of 64 percent. to have properties that did not have an exact match. We also assigned some services with uncertain properties. Upon the5 ROSSE EVALUATION completion of matching, the services in each list were returned with their match degrees in a descending order. ItTo evaluate ROSSE, we conducted a set of experiments. The should be noted that services with a match degree of zeroevaluation was focused on the accuracy and efficiency of were not returned.ROSSE in service discovery. In this section, we present the Let Rel be the set of relevant services, Ret be the set ofevaluation results. returned services, Retrel be the set of returned relevant5.1 Accuracy of ROSSE in Service Discovery services, RC represent recall, and P REC represent preci- sion. We defineROSSE can discover WSDL/UDDI and OWL-S services. Wecompare ROSSE with UDDI keyword matching and the jRe trelj jRe treljOWL-S matching [15], respectively, from the aspect of RC ¼ ; P REC ¼ : jRe lj jRe tjaccuracy in service discovery. First, we show how ROSSEincreases the similarity degrees of relevant services to a Figs. 5 and 6 show the averaged results of the 10 tests inservice query. group 1 and group 2, respectively. We observe that ROSSE achieves the best performance in5.1.1 Increased Similarity Degrees of ROSSE the tests of group 1, whereas UDDI has the worstTable 3 shows five services (S1-S5) that are relevant to a performance because of its keyword matching. For exam-service query. Each service has five properties (P1-P5), of ple, when the recall is 30 percent, the precisions of ROSSE,which P1 and P4 are dependent properties. Except for OWL-S, and UDDI are 95 percent, 84 percent, andservice S1, all the other four services S2-S5 have uncertain 65 percent, respectively. This is mainly because of theproperties in their service records. The match degrees of capability of ROSSE in the reduction of dependent proper-exact, plug-in, subsume, and uncertain are assigned 100 percent, ties and its use of semantic inference techniques, as outlined87 percent, 50 percent, and 50 percent, respectively. in Sections 2.2 and 2.3, respectively. We observe that UDDI keyword matching, OWL-S We observe that in most cases, ROSSE shows the bestmatching, and ROSSE produce different similarity degrees performance in the tests of group 2. However, OWL-Swhen matching the five services. In the case of service S1, performs better than ROSSE in two cases when recall isUDDI has a match degree of 40 percent, OWL-S has 10 percent and 30 percent, respectively. The reason is that85 percent, and ROSSE has 96 percent. UDDI produces the OWL-S does not deal with uncertain properties. As a result,
  9. 9. LI ET AL.: GRID SERVICE DISCOVERY WITH ROUGH SETS 859Fig. 5. The performance of ROSSE, OWL-S, and UDDI in the tests of Fig. 7. The performance of ROSSE in group 1 and group 1. and the OWL-S algorithm [15], respectively. RACER wassome irrelevant services are not returned because of the used by the OWL-S algorithm to reason the relationships ofexistence of uncertain properties. This leads to a high properties. As described in Section 4.1, RACER has a time-precision for some instances. We also observe that in the intensive initialization process for parsing every OWLtests of group 2, UDDI performs better than OWL-S in some document. We implemented a light-weighted reasoner incases. The reason is that some relevant services that have ROSSE to overcome the high overhead incurred by RACERuncertain properties are not matched by OWL-S. However, when parsing multiple OWL documents. We compared thethese relevant services are matched by UDDI because of the overhead of the ROSSE reasoner with that of RACER usingexistence of an exact match in their properties. two ontologies defined in two OWL documents, and the It should be noted that both OWL-S and UDDI do not results are shown in Fig. 8.reach a recall of 100 percent in the tests of group 2 because We observe that the difference in overhead betweenof their limitations in service matching. For example, OWL- RACER and the ROSSE reasoner gets larger with anS only matches six of the 10 relevant services, and UDDI increase in the number of queries. This is because for eachmatches eight relevant services. query received, RACER performs two initialization pro- We also compared the performance of ROSSE in the tests cesses, which consumes roughly 12 seconds. It takes theof group 1 and group 2, respectively. The results are plotted ROSSE reasoner about 3.4 seconds to load the twoin Fig. 7 showing that ROSSE performs better in group 1 ontologies. However, once loaded by the ROSSE reasoner,than in group 2. This can be explained by the existence of the two ontologies are maintained in memory for access.uncertain properties of services in the tests of group 2 in The overhead to parse a query is just a few hundredwhich ROSSE loses some useful information on relevant microseconds. As a result, the overhead of the ROSSEservices when matching the query. It is worth noting that reasoner does not change much with an increase in theROSSE matches all the relevant services in the tests of both number of 1 and group 2, and it performs reasonably well in To evaluate the efficiency of ROSSE in service discovery,terms of precision and recall. we registered 10,000 computing services with ROSSE. Each service had five properties, of which two were dependent5.2 Efficiency of ROSSE in Service Discovery properties. We posted a service query with two properties.ROSSE was evaluated on a Pentium IV 2.6-GHz machine Therefore, 10 ontology queries were parsed by the ROSSEwith 512 Mbytes of RAM running Red Hat Fedora Linux 3. reasoner and RACER, respectively, when matching theWe compared the efficiency of ROSSE with that of UDDI query. Service properties were defined in one ontology. WeFig. 6. The performance of ROSSE, OWL-S, and UDDI in the tests of Fig. 8. The overhead of the ROSSE reasoner in the access of twogroup 2. ontologies.
  10. 10. 860 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 20, NO. 6, JUNE 2008 Fig. 10. Efficiency of ROSSE in accessing service records.Fig. 9. The overhead of ROSSE in matching services. description models, and it takes into account robustness,performed two groups of tests. In group 1, we compared the efficiency, and security issues.overhead of ROSSE with that of UDDI and OWL-S, Semantic Web technologies [12] can be used to furtherrespectively, in matching services. In group 2, we evaluated enhance service discovery. As shown in Fig. 1, services canthe efficiency of ROSSE when accessing service records. be annotated with metadata whose relationships areFigs. 9 and 10 show the evaluation results of the two typically defined with a domain ontology. One key technol-groups, respectively. ogy to facilitate service discovery with semantic annotations In Fig. 9, we observe that UDDI has the least overhead is OWL-S, an OWL-based ontology for encoding propertieswhen matching services. This is because UDDI only of Web services. OWL-S ontology defines a service profilesupports keyword matching. It does not incur a reasoning for encoding a service description, a service model forprocess, which is usually time consuming. The overhead of specifying the behavior of a service, and a service groundingROSSE in matching services is mainly caused by the process for invoking the service. Typically, a service discoveryof reducing dependent properties. The reduction process process involves a matching between the profile of a servicegets slower with an increase in the number of services. We advertisement and the profile of a service request usingobserve that the overhead of OWL-S matching does not domain ontologies described in OWL. The service profilechange much with an increase in the number of services. describes not only the functional properties of a service suchThis is because the overhead of OWL-S matching is mainly as its inputs, outputs, preconditions, and effects (IOPEs) butcaused by the initialization process of RACER. Other also nonfunctional features such as name, category, andoverhead involved in OWL-S matching is small, e.g., the QoS-related aspects of a service. Srinivasan et al. [34]overhead of RACER in parsing an ontology query is just a enhanced UDDI for service discovery by embedding OWL-few hundred microseconds. As a result, the number of S in a UDDI registry. Paolucci et al. [15] present aservices involved does not make much change to the overall matchmaking algorithm for discovery of services withoverhead of OWL-S matching. OWL-S interfaces. Building on this algorithm, a number of In Fig. 10, we observe that ROSSE performs best when extensions are available. For example, Jaeger et al. [16]accessing service records due to its reduction of dependent introduce “contravariance” in matching inputs and outputsproperties. The OWL-S matching has a similar performance between service advertisements and service requests usingto UDDI in this process. OWL-S, Li and Horrocks [17] introduce an “intersection” relationship between a service advertisement and a service6 RELATED WORK request, and Majithia et al. [18] introduce reputation metrics in matching services.As the computational grid is evolving toward a service- Besides OWL-S, another prominent effort for semanticoriented computing infrastructure, service discovery has annotations of services is WSMO [19], which is built on fourbeen a research focus in the grid community. Grid key concepts—ontologies, standard Web services withinformation services such as Globus MDS [31] and WSDL interfaces, goals, and mediators. WSMO stressesR-GMA [32] facilitate discovery of resources and services the role of a mediator in order to support interoperationin a grid environment. However, they are restricted to between Web services. A mechanism is also proposed forkeyword-based queries. UDDI is an industry initiative for discovery of WSMO services [20].discovery of Web services. UDDI has been utilized by the Although the aforementioned approaches and algo-grid community for discovery of grid services [7], [8], [11]. rithms are available to facilitate the discovery of gridSimilar to Globus MDS, UDDI only supports keyword services, these efforts have never considered uncertainty ofmatching when searching for services. Various UDDI properties when matching services. As a result, they mayextensions have been proposed to enhance service dis- potentially produce a low accuracy when matching ser-covery [9], [10], [11]. Among them, UDDI-M [11] is flexible vices. Building on the Rough sets theory, ROSSE is capablein attaching metadata defined in RDF triples to various of reducing uncertain properties. In this way, ROSSEentities associated with a service. Building on UDDI, the increases the accuracy of service discovery. ROSSE repre-Grimoires service registry [33] supports multiple service sents an initial but significant advance toward solving
  11. 11. LI ET AL.: GRID SERVICE DISCOVERY WITH ROUGH SETS 861uncertainty of properties when matching services for high [5] K. Czajkowski, D.F. Ferguson, I. Foster, J. Frey, S. Graham, I. Sedukhin, D. Snelling, S. Tuecke, and W. Vambenepe, “The WS-accuracy in service discovery. Resource Framework,” wsrf.pdf, Mar. 2004. [6] B. Sotomayor and L. Childers, Globus Toolkit 4: Programming Java7 CONCLUSION AND FUTURE WORK Services. Morgan Kaufmann, 2005.In this paper, we have presented ROSSE, a search engine for [7] S. Banerjee, S. Basu, S. Garg, S. Garg, S.J. Lee, P. Mullan, and P. Sharma, “Scalable Grid Service Discovery Based on UDDI,” Proc.discovery of grid services. ROSSE builds on the Rough sets Third Int’l Workshop Middleware for Grid Computing (MGC ’05),theory to dynamically reduce uncertain properties when pp. 1-6, Dec. 2005.matching services. In this way, ROSSE increases the [8] B. Sinclair, A. Goscinski, and R. Dew, “Enhancing UDDI for Gridaccuracy of service discovery. The evaluation results have Service Discovery by Using Dynamic Parameters,” Proc. Int’l Conf. Computational Science and Its Applications (ICCSA ’05), pp. 49-59,shown that ROSSE significantly improves the precision and May 2005.recall compared with UDDI keyword matching and OWL-S [9] A. ShaikhAli, O.F. Rana, R.J. Al-Ali, and D.W. Walker, “UDDIe:matching, respectively. We have also introduced a QoS An Extended Registry for Web Service,” Proc. Symp. Applicationsmodel to filter functionally matched services with their and the Internet Workshops (SAINT ’03), pp. 85-89, Jan. 2003. [10] A. Powles and S. Krishnaswamy, “Extending UDDI withQoS-related nonfunctional performance. To maximize user Recommendations: An Association Analysis Approach,” Proc.satisfaction in service discovery, ROSSE dynamically Joint Workshop Web Services and Model-Driven Enterprise Informationdetermines the set of services that will be presented to Services (WSMDEIS ’05), pp. 45-54, May 2005.users based on the lower and upper approximations of [11] S. Miles, J. Papay, V. Dialani, M. Luck, K. Decker, T. Payne, and L. Moreau, “Personalised Grid Service Discovery,” IEE Proc. Soft-relevant services. We expect to carry out the following work ware, special issue on performance eng., vol. 150, no. 4, pp. 252-to improve ROSSE in the future: 256, 2003. [12] T. Berners-Lee, J. Hendler, and O. Lassila, “The Semantic Web,” . Efficiency. It has been shown that finding a minimal Scientific Am., vol. 284, no. 4, pp. 34-43, 2001. reduct in Rough sets is an NP-hard problem when [13] D.L. Martin, M. Paolucci, S.A. McIlraith, M.H. Burstein, D.V. the number of properties gets large [28]. Heuristic McDermott, D.L. McGuinness, B. Parsia, T.R. Payne, M. Sabou, M. Solanki, N. Srinivasan, and K.P. Sycara, “Bringing Semantics to methods need to be investigated to speed up the Web Services: The OWL-S Approach,” Proc. First Int’l Workshop process in service property reduction. Semantic Web Services and Web Process Composition (SWSWPC ’04), . Scalability. The number of services that are registered pp. 26-42, July 2004. with ROSSE could be large. Scalability is another [14] D.L. McGuinness and F. van Harmelen, OWL Web Ontology Language Overview, World Wide Web Consortium (W3C) recom- issue that needs to be addressed. UDDI version 3 mendation,, Feb. 2004. provides supports for multiple registries, but the [15] M. Paolucci, T. Kawamura, T. Payne, and K. Sycara, “Semantic specification does not specify how these registries Matching of Web Service Capabilities,” Proc. First Int’l Semantic should be structured to enhance scalability in service Web Conf. (ISWC ’02), pp. 333-347, June 2002. registration. Distributed Hash Table (DHT)-based [16] ¨ M.C. Jaeger, G. Rojec-Goldmann, G. Muhl, C. Liebetruth, and K. Peer-to-Peer (P2P) systems such as Chord [29] and Geihs, “Ranked Matching for Service Descriptions Using OWL-S,” Proc. Comm. in Distributed Systems (KiVS ’05), pp. 91-102, Feb. 2005. Pastry [30] have shown their efficiency and scal- [17] L. Li and I. Horrocks, “A Software Framework for Matchmaking ability in content distribution and lookup. We expect Based on Semantic Web Technology,” Int’l J. Electronic Commerce, that ROSSE’s scalability in service discovery can be vol. 8, no. 4, pp. 39-60, 2004. improved with DHT-structured P2P systems. [18] S. Majithia, A.S. Ali, O.F. Rana, and D.W. Walker, “Reputation- . Deployment. Services, once discovered by ROSSE and Based Semantic Service Discovery,” Proc. 13th IEEE Int’l Workshops Enabling Technologies: Infrastructures for Collaborative Enterprises selected by the user, should be dynamically deployed (WETICE ’04), pp. 297-302, 2004. and invoked [43]. Such deployment could be made [19] D. Roman, U. Keller, H. Lausen, J. de Bruijn, R. Lara, M. Stollberg, part of an existing workflow engine such as Triana A. Polleres, C. Feier, C. Bussler, and D. Fensel, “Web Service [47], which could utilize such a discovery technique Modeling Ontology,” Applied Ontology, vol. 1, no. 1, pp. 77-106, like ROSSE as part of the enactment process. 2005. [20] U. Keller, R. Lara, A. Polleres, I. Toma, M. Kifer, and D. Fensel, . Ontology alignment [44]. Services may be advertised “WSMO Discovery,” Working Draft D5.1v0.1, WSMO, http:// with properties that follow distinct ontologies. To, 2004. further increase the accuracy of service discovery, [21] M. Li and M.A. Baker, The Grid: Core Technologies. John Wiley ROSSE needs to be enhanced with the existing Sons, 2005. efforts on ontology alignment such as [45] and [46]. [22] M. Li, P. van Santen, D.W. Walker, O.F. Rana, and M.A. Baker, “SGrid: A Service-Oriented Model for the Semantic Grid,” Future Generation Computer Systems, vol. 20, no. 1, pp. 7-18, 2004.REFERENCES [23] M. Li, B. Yu, C. Huang, and Y.H. Song, “Service Matchmaking[1] F. Curbera, M. Duftler, R. Khalaf, W. Nagy, N. Mukhi, and S. with Rough Sets,” Proc. Sixth IEEE Int’l Symp. Cluster Computing Weerawarana, “Unraveling the Web Services: An Introduction to and the Grid (CCGrid ’06), pp. 23-30, May 2006. SOAP, WSDL, and UDDI,” IEEE Internet Computing, vol. 6, no. 2, [24] B. Yu, W. Guo, M. Li, Y.H. Song, P. Hobson, and M. Qi, “Service pp. 86-93, 2002. Matchmaking and Discovery with Rough Sets,” Proc. Second Int’l[2] I. Foster and C. Kesselman, The Grid, Blueprint for a New Computing Conf. Semantics, Knowledge and Grid (SKG ’06), p. 80, Nov. 2006. Infrastructure. Morgan Kaufmann, 1998. [25] Z. Pawlak, “Rough Sets,” Int’l J. Computer and Information Science,[3] M.P. Atkinson, D. De Roure, A.N. Dunlop, G. Fox, P. Henderson, vol. 11, no. 5, pp. 341-356, 1982. A.J.G. Hey, N.W. Paton, S. Newhouse, S. Parastatidis, A.E. [26] L. Zeng, B. Benatallah, A.H.H. Ngu, M. Dumas, J. Kalagnanam, Trefethen, P. Watson, and J. Webber, “Web Service Grids: An and H. Chang, “QoS-Aware Middleware for Web Services Evolutionary Approach,” Concurrency—Practice and Experience, Composition,” IEEE Trans. Software Eng., vol. 30, no. 5, pp. 311- vol. 17, no. 2-4, pp. 377-389, 2005. 327, May 2004.[4] I. Foster, C. Kesselman, J.M. Nick, and S. Tuecke, “Grid Services [27] ¨ V. Haarslev and R. Moller, “Description of the RACER System for Distributed System Integration,” Computer, vol. 35, no. 6, and Its Applications,” Proc. Int’l Workshop Description Logics (DL pp. 37-46, 2002. ’01), Aug. 2001.
  12. 12. 862 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 20, NO. 6, JUNE 2008[28] A. Skowron and C. Rauszer, “The Discernibility Matrices and Maozhen Li received the PhD degree from the Functions in Information Systems,” Decision Support by Experi- Institute of Software, Chinese Academy of ence—Application of the Rough Sets Theory, R. Slowinski, ed. Kluwer Sciences, Beijing, in 1997. He is a lecturer in Academic Publishers, pp. 331-362, 1992. the School of Engineering and Design, Brunel[29] I. Stoica, R. Morris, D. Liben-Nowell, D. Karger, M. Kaashoek, F. University. His research interests are in the Dabek, and H. Balakrishnan, “Chord: A Scalable Peer-to-Peer areas of grid computing, distributed problem- Lookup Protocol for Internet Applications,” IEEE Trans. Networks, solving environments for large-scale simula- vol. 11, no. 1, pp. 17-32, 2003. tions, intelligent systems, service-oriented com-[30] A. Rowstron and P. Druschel, “Pastry: Scalable, Distributed Object puting, and semantic Web. He has more than Location and Routing for Large-Scale Peer-to-Peer Systems,” Proc. 50 publications in these areas. He coauthored 18th IFIP/ACM Int’l Conf. Distributed Systems Platforms (Middleware The Grid: Core Technologies, a research-level textbook on grid ’01), pp. 329-350, Nov. 2001. computing published by John Wiley Sons in 2005. He is on the[31] J.M. Schopf, L. Pearlman, N. Miller, C. Kesselman, I. Foster, M. editorial boards of the Encyclopedia of Grid Computing Technologies D’Arcy, and A. Chervenak, “Monitoring the Grid with the Globus and Applications and the International Journal of Grid and High Toolkit MDS4,” J. Physics: Conf. Series, vol. 46, pp. 521-525, 2006. Performance Computing. He has been serving as a TPC member for[32] A.W. Cooke et al., “The Relational Grid Monitoring Architecture: various conferences in the area of grid computing, e.g., IEEE CCGrid Mediating Information about the Grid,” J. Grid Computing, vol. 2, 2005, CCGrid 2006, CCGrid 2007, CCGrid 2008, IEEE SKG 2005, SKG no. 4, pp. 323-339, 2004. 2006, SKG 2007, and IEEE CSE 2008. He is a member of the IEEE.[33] W. Fang, S. Miles, and L. Moreau, “Performance Analysis of a Semantics-Enabled Service Registry,” Concurrency and Computa- Bin Yu received the PhD degree from the tion: Practice and Experience, vol. 20, no. 3, pp. 207-223, 2008. School of Engineering and Design, Brunel[34] N. Srinivasan, M. Paolucci, and K.P. Sycara, “An Efficient University, in April 2007. He is currently a Algorithm for OWL-S Based Semantic Search in UDDI,” Proc. system analyst at Levele Ltd., Edinburgh. His First Int’l Workshop Semantic Web Services and Web Process research interests are in the areas of service- Composition (SWSWPC ’04), pp. 96-110, July 2004. oriented computing, grid computing and ap-[35] N. Zhong and A. Skowron, “A Rough Set Based Knowledge plications, service discovery, and composition Discovery Process,” Int’l J. Applied Math. and Computer Science, optimization. vol. 11, no. 3, pp. 603-619, 2001.[36] C. Caceres, A. Fernandez, S. Ossowski, and M. Vasirani, “Agent- Based Semantic Service Discovery for Healthcare: An Organiza- tional Approach,” IEEE Intelligent Systems, vol. 21, no. 6, pp. 11-20, 2006. Omer Rana received the PhD degree in neural[37] V. Tsetsos, C. Anagnostopoulos, and S. Hadjiefthymiades, “On the computing and parallel architectures from the Evaluation of Semantic Web Service Matchmaking Systems,” Proc. Imperial College of Science, Technology, and Fourth IEEE European Conf. Web Services (ECOWS ’06), pp. 255-264, Medicine, London University. He is a professor Dec. 2006. of performance engineering in the School of[38] D.A. Buell and D.H. Kraft, “Performance Measurement in a Fuzzy Computer Science and the deputy director of the Retrieval Environment,” Proc. ACM SIGIR ’81, pp. 56-62, 1981. Welsh eScience Center at Cardiff University,[39] C. van Rijsbergen, Information Retrieval. Butterworths, 1979. United Kingdom. His research interests include[40] K. Lee, J. Jeon, W. Lee, S. Jeong, and S. Park, QoS for Web Services: high-performance distributed computing, multia- Requirements and Possible Approaches, World Wide Web Consor- gent systems, and data mining. tium (W3C) Working Group Note 25, 2003.[41] W. Smith, I. Foster, and V. Taylor, “Predicting Application Run Times Using Historical Information,” Proc. Fourth Workshop Job Zidong Wang received the PhD degree in Scheduling Strategies for Parallel Processing, pp. 122-142, 1998.[42] L. Gong, X. Sun, and E.F. Watson, “Performance Modeling and electrical and computer engineering from the Prediction of Nondedicated Network Computing,” IEEE Trans. Nanjing University of Science and Technology, Computers, vol. 51, no. 9, pp. 1041-1055, Sept. 2002. Nanjing, China, in 1994. He is a professor of[43] L. Qi, H. Jin, I.T. Foster, and J. Gawor, “HAND: Highly Available dynamical systems and computing at Brunel Dynamic Deployment Infrastructure for Globus Toolkit 4,” Proc. University. His research interests include dyna- mical systems, signal processing, bioinfor- 15th Euromicro Int’l Conf. Parallel, Distributed and Network-Based matics, and control theory and applications. He Processing (PDP ’07), pp. 155-162, 2007.[44] N.F. Noy, “Semantic Integration: A Survey of Ontology-Based has published more than 80 papers in refereed Approaches,” ACM SIGMOD Record, vol. 33, no. 4, pp. 65-70, 2004. international journals. He is currently serving as[45] B. Kryza, R. Slota, M. Majewska, J. Pieczykolan, and J. Kitowski, an associate editor for the IEEE Transactions on Automatic Control, the “Grid Organizational Memory—Provision of a High-Level Grid IEEE Transactions on Signal Processing, the IEEE Transactions on Systems, Man, and Cybernetics—Part C, the IEEE Transactions on Abstraction Layer Supported by Ontology Alignment,” Future Control Systems Technology, and Circuits, Systems Signal Proces- Generation Computer Systems, vol. 23, no. 3, pp. 348-358, 2007.[46] R. Pan, Z. Ding, Y. Yu, and Y. Peng, “A Bayesian Network sing, an action editor for Neural Networks, an editorial board member for Approach to Ontology Mapping,” Proc. Int’l Semantic Web Conf. the International Journal of Systems Science, Neurocomputing, the (ISWC ’05), pp. 563-577, 2005. International Journal of Computer Mathematics, and the International[47] I. Taylor, M. Shields, I. Wang, and A. Harrison, “Visual Grid Journal of General Systems, and an associate editor on the conference Workflow in Triana,” J. Grid Computing, vol. 3, nos. 3-4, pp. 153- editorial board for the IEEE Control Systems Society. He is a senior member of the IEEE, a fellow of the Royal Statistical Society, a member 169, 2005. of the program committee for many international conferences, and a very active reviewer for many international journals. He was nominated as an appreciated reviewer for IEEE Transactions on Signal Processing in 2006 and an outstanding reviewer for IEEE Transactions on Automatic Control in 2004 and for the journal Automatica in 2000. . For more information on this or any other computing topic, please visit our Digital Library at