1 WSPDS: Web Services Peer-to-peer Discovery Service† Farnoush Banaei-Kashani, Ching-Chien Chen, and Cyrus Shahabi Computer Science Department, University of Southern California, Los Angeles, California 90089 [banaeika,chingchc,shahabi]@usc.edu
Abstract— The Web Services infrastructure is a dis- One could anticipate popularity of this infrastruc- tributed computing environment for service-sharing. In this ture in advance, because it is an extension of the environment, resource discovery is required as a primitive functionality for users to be able to locate the services, the successful browser-based web-programming tech- shared resources. A discovery service with centralized ar- nology to a general distributed application devel- chitecture, such as UDDI, restricts the scalability of this opment environment. However, more importantly, environment as it grows to the scales comparable with the success of this infrastructure must be attributed to size of the web itself. In addition, current extensively used web service standards (e.g. UDDI, WSDL), do not support its fundamental features: discovery at a semantic level. In this paper, we introduce • Loose coupling: services developed and deployed WSPDS (Web Services Peer-to-peer Discovery Service), a independently using heterogeneous platforms can fully decentralized and interoperable discovery service with semantic-level matching capability. We believe the peer-to- be integrated seamlessly to build distributed appli- peer architecture of the semantic-enabled WSPDS not only cations with new functionalities; hence, interoper- satisfies the design requirements for efficient and accurate ability. Loose coupling is mainly enabled by XML- discovery in distributed environments, but also is compati- based SOAP communication specification, which ble with the nature of the Web Services environment as a self-organized federations of peer service-providers without allows platform-independent information exchange any particular sponsor. between services. Keywords— Web Services discovery, Peer-to-peer • Full decentralization: all communications of the discovery, Ontology, Semantic matching interacting entities are in a peer-to-peer fashion, without any central coordination; hence, scalabil- I. Introduction ity. The Web Services programming infrastructure • Semantic level search: this feature allows web is the current generation of a succession of sys- service requesters to search for published web ser- tems proposed to develop distributed applications: vices not only based on keywords, but also based RPC, CORBA, DCOM, and now Web Services. A on ontological concepts. web service is a self-contained application module A. Discovery Service for Web Services with well-described functionality that can be in- voked across the web. The Web Services program- In general, in a distributed computing system a ming environment is a distributed computing envi- discovery service locates (or discovers) resources ronment in which participants share their services; dispersed across the system in response to re- hence, a service-sharing environment. Each partic- source discovery queries issued by the system enti- ipant can potentially act both as a service provider ties. With Web Services, resources are the services and as a client. As a service provider, the partic- shared on the web. To be specific, a discovery ser- ipant builds and optionally shares its services for vice for Web Services is itself a web service that public use. As a client, on the other hand, the locates the service description document(s) of the participant can develop distributed applications by service(s) that hit a service query. A service de- discovery and seamless integration of the public ser- scription document (e.g., a WSDL file) provides vices with its own private services. both abstract and concrete information required The Web Services infrastructure is adopted more for proper invocation of a service. A service query rapidly and widely as compared to its predecessors. characterizes a set of services with particular char- acteristics, such as name, abstract(or description), † This research has been funded in part by NSF grants interface model, etc., to be located. EEC-9529152 (IMSC ERC), IIS-0082826 (ITR), IIS-0238560 (CAREER), IIS-0324955 (ITR) and IIS-0307908, and unre- B. Design Issues and Approaches for Discovery stricted cash gifts from Okawa Foundation and Microsoft. Any opinions, findings, and conclusions or recommendations Service expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science To be compatible with the fundamental features Foundation. of the Web Services infrastructure (as discussed ion. Entities are peers in functionality and each entity is potentially both a server and a client of the peer-to-peer service; hence, sometimes entities with User with
Communication are referred to as servents (i.e., server and client). The Web Services discovery service can be im- Communication User Query User t Communication with Neighbor n Engine Neighbor Query e v r Response to Neighbor Query e
Response to User Query User to Response plemented as a peer-to-peer service, eliminating de- S S D P S W pendency on a distinct service provider. Each ser-
A
Local Query vent serves others by providing information about Engine its own web services in response to queries, and in turn, as a client it issues discovery queries to locate the web services that are not available locally. Ser- vents build a network in which each servent has a Local Inspection Documents few other servents as neighbors. When a servent re- Fig. 1. WSPDS Architecture ceives a request for a web service from the local user and cannot find the web service locally, as a client it originates a discovery query and propagates the re- quest into the network through its neighbors. Ser- above), a discovery service should support the fol- vents collaborate based on a distributed algorithm lowing requirements: to disseminate the query. During propagation of • Interoperablity, to be integrable with other web the query, if a servent finds the requested web ser- services, to support different service description vice locally, it responds to the originator by pro- standards, and to be portable to different plat- viding its location and description. forms; Additionally, in order to achieve efficient query • Scalability, to grow to the web scales without be- propagation in a peer-to-peer environment, the ing a performance bottleneck; linkage between servents should be built based on • Efficiency, to support the dynamic environment the hosted data contents (e.g., web service descrip- of the Web Services with frequent changes/updates tions) of the servents. Finally, a more accurate of the location of the services and their description match will be accomplished by annotating both the documents; advertised web services and users’ requests with • Fault tolerance, to be resistant to unwanted globally shared concepts. breakdowns and malicious attacks. • Semantic based discovery, to find a match based II. Peer-to-peer Discovery Service for on the common conceptual space of service re- Web Services questers and providers. The Web Services infrastructure is a self- We argue that as compared to a centralized ar- organized federation of service providers for service- chitecture (e.g., UDDI [1], the currently used stan- sharing. Thus, a peer-to-peer architecture is an dard for globally publishing and locating web ser- appropriate choice for the discovery service in this vices), a decentralized design for the Web Services environment. Considering the usual autonomous discovery service is more scalable (obviously), more behavior of the service providers, an unstructured fault tolerant (by eliminating the single point of peer-to-peer discovery service is preferred. Here, failure), and more efficient (by reducing the over- we introduce WSPDS (Web Services Peer-to-peer head of centralized update of the discovery service). Discovery Service), a fully decentralized and in- Distributed directory services and peer-to-peer ser- teroperable discovery service with an unstructured vices are two alternative service models with decen- peer-to-peer architecture. tralized architecture. Distributed directory servers are usually dedicated facilities that are built and A. Architecture maintained under unique management to provide service to the clients of a distributed environment. WSPDS is a distributed discovery service im- However, the Web Services infrastructure is a self- plemented as a cooperative service. A network organized federation of peer entities without any of WSPDS servents collaborate to resolve discov- particular sponsor for the system. It is desirable ery queries raised by their peers. Figure 1 depicts that the federation lives, changes, and expands an unstructured peer-to-peer network of WSPDS independent of any distinct service facility with servents. Each servent is composed of two en- global authority. With peer-to-peer services, the gines, communication engine and local query en- role of distinct service providers is eliminated. Sys- gine, standing for the two roles that a servent plays: tem entities all cooperate to provide a service as a 1. Communication and Collaboration: the commu- result of group collaboration in a distributed fash- nication engine provides the interface to user and also represents the servent in the peer-to-peer net- erative discovery. These mechanisms are mostly work of servents. This engine is responsible for the compatible with the Gnutella peer-to-peer network following tasks: specification [2] enhanced by our novel technique termed probabilistic flooding. In [3], we prove that • Receiving service queries from users, resolving this technique improves scalability of Gnutella’s the queries by local query (through the local flooding-based dissemination mechanism by up to query engine) and global query (via its peer 99%, effectively eliminating the major drawback of servents), and finally merging the received re- this Gnutella-like peer-to-peer discovery system. sponses to reply to the user query; and • Receiving queries from its neighbors in the peer- A.1 Network Setup to-peer network, resolving the queries by local Each servent maintains a list of the most recently query, and sending the response (if not empty) active servents of the network, denoted as servent to the network as well as forwarding the query cache. Each time a servent is re-activated, it probes (if query has still some time to live, i.e., TTL > the servents listed in the servent cache to find k 0) to other neighbors in the network. nodes that are still active and designates them as 2. Local query: the local query engine receives the its neighbors. In this way, a new servent can join queries from the communication engine, queries the the peer-to-peer network based on the local infor- local site (where the servent is running) for match- mation without any unique global control. For the ing services, and sends responses to the communi- first time a servent is activated, the servent cache cation engine. contains access points of a few WSPDS servents In the following sections, we first explain the im- associated with some large service providers that plementation of the two engines to build a primitive are almost always active. When a servent joins the WSPDS network based on the basic peer-to-peer network, it periodically uses a Gnutella-like ping- network specification Gnutella[2]. The primitive pong mechanism to find other active servents in the WSPDS supports only keyword-matching queries. network and refreshes its local servent cache to be Thereafter, we describe our approach to add onto- updated for the next re-activation. logical concepts to the primitive WSPDS to achieve semantic-based peer-to-peer network construction A.2 Cooperative Discovery (termed Sem-WSPDS) and service discovery. To discover a service requested by user, a servent III. Construction of a Primitive originates a query (enveloped in a SOAP message) Peer-to-Peer Network of WSPDS in the network of servents. The servents collaborate Servents to propagate the query based on the probabilistic- flooding dissemination mechanism. Dissemination A. Communication Engine of a query is restricted by its TTL. A servent that Consider to build a peer-to-peer network of receives a copy of the query message decreases TTL WSPDS servents based on Gnutella protocol. of the query by 1, and if TTL > 0, forwards the The communication engine of a WSPDS servent query to each of its neighbors with the probability exchanges SOAP-enveloped query/response mes- p (p is in the interval [0.01, 0.1]). sages with 1) user applications/services, or 2) Besides forwarding the query messages, when a other WSPDS servents. The only difference be- servent receives a query it also inspects the local tween these two types of communications is a site for matching services. If the local inspection unique identifier and a TTL field embedded in the results in discovering one or more services, the ser- MessageDescriptor of the messages exchanged be- vent prepares a response message and sends it back tween two servents (with the second case above), towards the originator of the query. The response for peer-to-peer collaboration purposes. Obviously, message traverses the path of the query message these fields are not required for the messages com- in the reverse order. To enable returning the re- municated between a user application and WSPDS sponse messages to the originator, a query origi- servent (the first case). Figure 2 shows sample com- nator marks its query message by a unique iden- munications between a WSPDS servent and a user. tifier. The servents in the path of a query cache Figure 5 (see the appendix) depicts the main the identifiers of the query in a short-lived buffer. routine that implements the communication and When they receive a response message, they match message handling tasks of the communication en- the identifier of the response message (which is the gine. Instead, here we focus on the mechanisms same as the identifier of the corresponding query) implemented by the communication engines of the against the buffered identifiers and forward the re- peer servents to 1) build and maintain the peer- sponse message to the neighbor from which they to-peer network of servents, and 2) execute coop- have received the corresponding query.
POST /WSPDS.asmx HTTP/1.1 HTTP/1.1 200 OK Host: micron34.usc.edu Content-Type: text/xml; charset=utf-8 Content-Type: text/xml; charset=utf-8 Content-Length: length Content-Length: length SOAPAction: http://micron34.usc.edu/SearchService
B. Local Query Engine query dissemination. With probabilistic flooding, this overhead is significantly reduced. However, as WSPDS queries allow keyword-matching queries we illustrate in the following sections, we believe on service name/abstarct, provider name and that a content-based peer-to-peer network, such as tModel[1](see Figure 2-a). These are the most com- QDN [5], can further reduce the overhead. Second, mon features currently used with discovery directo- keyword-matching is insufficient for discovering de- ries (e.g., UDDI) and service inspection documents sired web services, because it ignores semantic cor- (e.g., WSIL documents [4]) to characterize a ser- respondences. Since web service advertisers and vice. This set of query features is extendible to requesters may look at the same service from dif- support future interesting features (e.g. QoS) of ferent perspectives and express the service identity the service. in different ways, a discovery service should rely on The query engine applies the WSIL specification the semantic information to evaluate the similarity to inspect the local site and find services matching between the query and the advertised web services. (string-based) with the received query. A WSIL document (e.g., Figure 3-a) lists references to the A. Semantic-annotated Web Service Description description documents (e.g., WSDL) and (possibly) There have been a number of efforts to add se- UDDI records for the services available at the local mantics to web service description. Ontology has site. For each service, the WSIL file also provides been identified as the basis for semantic annotation. some metadata, such as web service name. The An ontology specifies shared expressions of con- local query engine of the WSPDS servent parses cepts and agreements on the terminology/meaning the WSIL document of the local site and matches for communication. DAML-S profile module [6] the query against the metadata in the WSIL doc- and semantic-annotated WSDL [7] are two emerg- ument itself, as well as the metadata in the refer- ing web service descriptions based on ontology. enced service description documents and directory DAML-S profile module is a DAML+OIL ontol- records. Pointers to the locations of the WSDL for ogy for describing web services by defining “what the matching services are included in the response a service does”. It can be used for discovery at the message. Due to the extensibility of WSIL specifi- semantic level. Semantic-annotated WSDL is an cation, the query engine of the WSPDS servent can XML-formatted web service description document support future service description specifications. based on WSDL, and is extended with DAML+OIL ontologies for the purpose of representing WSDL in IV. Construction of a Semantic-enabled a machine interpretable form like DAML-S profile Peer-to-Peer network (Sem-WSPDS) module. Both DAML-S and semantic-annotated of WSPDS Servents WSDL techniques can be utilized to add ontologies There are two major drawbacks with the primi- to web service descriptions and accomplish auto- tive WSPDS network described in previous section. mated semantic web services discovery. Our discov- First, as compared to the centralized architectures, ery service relies on the use of semantic-annotated the architecture of WSPDS has higher overhead of WSDL to describe web services interfaces, because WSDL has been accepted as the industry standard are multiple web services with various operations for web service description and most of the existing on the same node, we map each web service oper- web services support WSDL standards. In addi- ation to a virtual node and build the QDN based tion, WSDL provides communication level details on the virtual nodes. of web services and numerous tools are developed based on WSDL. WSIL and semantic-annotated C. Communication Engine WSDL can provide the same capability as DAML- The communication engine of a WSPDS servent S profile module without adding significant com- exchanges SOAP-enveloped query/response mes- plexity to the basic standards. Currently, both sages with 1) user applications/services, or 2) other DAML-S and semantic-annotated WSDL only ap- WSPDS servents. These messages are annotated ply ontologies on the operational interfaces (i.e. in- with ontologies (see Figure 4 for example). put and output parameters of the operations of the web services), not on the web service names or de- C.1 Network Setup scriptions. In this paper, we consider the semantic- matching on the operational interfaces only. During the network setup phase, the linkages be- Figure 6 (see the appendix) shows the descrip- tween nodes are constructed based on the data con- tion of a GeoService web service, which finds the tents of the servents. A newly added node n joins geographical areas, such as city, country and is- the QDN by linking to some other nodes in a range land, located at a given latitude. The WSDL file geographically close to n. To select the neighbors, utilizes an approach similar to that of Sivashan- the new node applies a semantic matching function mugam et al. [7] to annotate the input/output to evaluate the similarity between its input/output parameters of operations (e.g. getLocByLat) with and those of the other nodes, respectively. The new ontology (see Figure 3-b). The input Latitude is node links to the nodes that have more similar in- restricted to the concept Latitude as defined in put/output. The semantic matching function re- the coord.daml ontology, while the output is anno- lies on the MatchMaker algorithm proposed in [8] tated with the concept GeographicalArea defined to compute the semantic similarity. MatchMaker in geo-ont.daml ontology. The GeoService’s WSIL utilizes DAML+OIL logic to infer the similarity. file stored in the registry is shown in Figure 3-a. C.2 Cooperative Discovery The service name/abstract can be queried directly from the WSIL, while input/output parameters for To discover a requested service, a SOAP- each operation can be retrieved by tracing the “de- enveloped query is originated at a servent in the scription:location” pointer of WSIL to a semantic- network (see Figure 4-a). Each servent that re- annotated WSDL. A possible user query is illus- ceives the query forwards it to the neighbor that trated in Figure 4-a. The query searches for ser- has the most similar identity to the query (again, vice(s) that accept instances of Latitude as input, we use MatchMaker to calculate the similarity). In and generate instances of City as output. addition to forwarding the query messages, when a servent receives a query it also inspects WSIL and B. Querical Data Network (QDN) the semantic-annotated WSDL (whose location is A QDN is a federation of a dynamic set of specified in WSIL) on the local site for matching lo- peer, autonomous nodes communicating through a cal services based on input and output ontologies. transient-topology interconnection. An identity for If the local inspection results in discovering one or each QDN node is defined based on its data con- more services, the servent prepares a response mes- tent. A node joins the QDN by linking to some sage and sends it back towards the originator of the other QDN nodes, selecting the nodes of “simi- query. The response message traverses the path of lar” identity with higher probability. The nodes the query message in the reverse order. who know the identity of their neighbors, route the D. Local Query Engine query to the neighbor that has the most similar identity to the target content (see [5] for more de- WSPDS queries allow semantic-matching queries tails about QDN). To illustrate how to build QDN on the service operational interfaces. It is ex- connections between WSPDS servents and how to tendible to support future interesting features (if perform capability matching between the web ser- semantic-enabled) such as service categories and vices on the QDN, in the following sections, we QoS. The query engine applies the WSIL specifi- consider a rather simple scenario where each node cation and the semantic-annotated WSDL to in- registers only one web service with one operation. spect the local site and find services matching with Under such circumstance, the identify of each ser- the received query. Pointers to the locations of the vent is defined as the ontologies associated with the WSDL for the matching services are included in the input/output parameters. For the case where there response message. The match between the service SOAPAction: http://micron34.usc.edu/SearchService
and the request is performed by comparing their in- This service is currently available online. We are put and output ontologies. The semantic-matching in the process of improving the primitive imple- also uses the main idea of MatchMaker algorithm; mentation based on the two concepts of content- i.e., the outputs of the query should be subsumed based peer-to-peer computing and ontology-based by the outputs of the service provided. Moreover, matching. We have already developed major com- if inputs of the query subsume the inputs of the ponents (i.e., semantic-matching and QDN-linking service, MatchMaker ranks the provided services routines) of the enhanced WSPDS, and expect to based on their input matching. For example, con- publish it for public use in near future. sider the advertised GeoService web service shown in Figure 6 and the query shown in Figure 4-a. References Their inputs match exactly, because they are re- [1] UDDI.org, “UDDI: Universal Description, Discovery and Integration of web services,” 2002, http://www.uddi.org/. stricted to the same ontological concepts (i.e. Lat- [2] Gnutella, “Gnutella RFC,” 2002, http://rfc- itude). Their outputs are also matching, since the gnutella.sourceforge.net/. [3] F. Banaei-Kashani and Cyrus Shahabi, “Criticality-based query concept City is a subclass of the service con- analysis and design of unstructured peer-to-peer networks cept GeographicalArea. Therefore, the web service as complex systems,” in Third International Workshop on Global and Peer-to-Peer Computing (GP2PC) in con- that is able to answer the geographical areas lo- junction with CCGrid’03, May 2003. cated at a given Latitude, commits to provide the [4] K. Ballinger, P. Brittenham, A. Malhotra, W.A. Nagy, and S. Pharies, “Specification: Web Services In- cities at the specified latitude. spection Language (WS-Inspection) 1.0,” November 2001, http://www.ibm.com/developerworks/library/ws- wsilspec.html. V. Related Work [5] F. Banaei-Kashani and C. Shahabi, “Searchable Quer- ical Data Networks,” in Proceedings of the Interna- A significant amount of recent research on web tional Workshop on Databases,Information Systems and services has focused on dynamic and automated Peer-to-Peer Computing in conjunction with VLDB’03, September 2003. web service composition [9, 10]. Towards this end, [6] DAML-S Coalition, “DAML-S: Web Service Description a vital step is to automatically and accurately dis- for the Semantic Web,” in Proceedings of the First Inter- national Semantic Web Conference, 2002. cover the web services with desired capabilities. [7] K. Sivashanmugam, K. Verma, A. Sheth, and J. Miller, The idea of using peer-to-peer (P2P) and ontol- “Adding Semantics to Web Services Standards,” in Pro- ceedings of the International Conference on Web Ser- ogy to discovery web services has been proposed by vices, 2003. [11, 12]. The P2P network utilized in our system [8] M. Paolucci, T. Kawmura, T. Payne, and K. Sycara, “Se- mantic Matching of Web Services Capabilities,” in Pro- is content-based and has a different architecture as ceedings of the First International Semantic Web Con- compared to that of [11]. In addition, our approach ference, 2002. [9] J. Cardoso and A. Sheth, “Semantic e-Workflow Compo- is different from [12] both on the architecture of sition,” Journal of Intelligent Information Systems, vol. P2P network and the utilization of semantic en- 21, no. 3, pp. 191–225, November 2003. [10] S. Ghandeharizadeh, C. Knoblock, C. Papadopoulos, abled web service description document. Another C. Shahabi, E. Alwagait, J. L. ambite, M. Cai, C.-C. Chen, feature that differentiate our system from theirs is P. Pol, R. Schmidt, S. Song, S. Thakkar, and R. Zhou, “Proteus: A System for Dynamically Composing and In- that all messages exchanged among WSPDS ser- telligently Executing Web Services,” in Proceedings of the vent are enveloped in SOAP. International Conference on Web Services, 2003. [11] K. Verma, K. Sivashanmugam, A. Sheth, A. Patil, S. Ound- hakar, and J. Miller, “METEOR-S WSDI: A Scalable In- VI. Conclusion frastructure of Registries for Semantic Publication and Dis- covery of Web Services,” Journal of Information Technol- We developed WSPDS that is a decentralized ogy and Management, under review. discovery service with peer-to-peer architecture for [12] M. Paolucci, K. P. Sycara, T. Nishimura, and N. Srini- vasan, “Using DAML-S for P2P Discovery,” in Proceedings the Web Services infrastructure. The primitive of the International Conference on Web Services, 2003. prototype of WSPDS is based on a variation of the Gnutella peer-to-peer network and keyword- Appendix matching between the web service descriptions. if (message is received from user) { //message is a user query forward the query to the local query engine; forward the query to all neighbors; } else //message is received from a neighboring servent; switch (MessageDescriptor) { case “RESPONSE”: ID=decodeDescriptor(MessageDescriptor); if (ID is one of my descriptor IDs) { merge the Result (from response) into the MergedResult with the same ID; if (time to respond to the user query is over) return the MergedResult to user; } else if (ID is in my routing table) forward the message according to the corresponding routing table entry; case “QUERY”: (ID,TTL)=decodeDescriptor(MessageDescriptor); add ID to the routing table; send the query to the local query engine; if (any matching service is found) respond to the query; if (TTL > 0) forward the query to each neighbor (except the sender) with probability ‘p’; case “PONG”: ID=decodeDescriptor(MessageDescriptor); if (ID is one of my descriptor IDs) add RespondingHostAddress to the servent cache ; else if (ID is in my routing table) route the pong message according to the corresponding routing table entry; case “PING”: (ID,TTL)=decodeDescriptor(MessageDescriptor); add ID to the routing table; if (local resources are sufficient for accepting a new neighbor) respond with pong; if (TTL > 0) forward the ping message to all neighbors (except the sender); }
Fig. 5. Message Processing at the Communication Engine (based on Gnutella) of a WSPDS Servent
Fig. 6. A WSDL document (geoservice.wsdl) annotated with ontologies(geo-ont.daml and coord.daml)