and sketch query evaluation on top of RSSDB. storage and querying. Specically , we introduce a formal whilst it outperforms, both in storage volumes and query vices in various user communities, lead nowadays to large facilitating the creation and exchange of metadata as any tiple schemas. Next, we present the design of a persistent net. The Resource Description Framework (RDF) aims at form of triples. Last, we briey presen t RQL, a declarative this paper, we advocate the use of database technology to edge. Our approach preserves the exibilit y of RDF in revolumes of RDF metadata. Managing such RDF resource support declarative access, as well as, logical and physical descriptions and schemas with existing low-level APIs and data model for RDF description bases created using mulMetadata are widely used in order to fully exploit informaindependence for voluminous RDF description bases. mation resources and the proliferation of description serexecution time, other approaches using a monolithic table tion resources available on corporate intranets or the Interother Web data. The growing number of available inforand easy maintenance of real-scale RDF applications. In ORDBMS by exploring the available RDF schema knowlRDF Store (RSSDB) for loading resource descriptions in an We present RDFSuite, a suite of tools for RDF validation, ning schemas and/or enriching descriptions at any time, le-based implemen tations does not ensure fast deployment language for querying both RDF descriptions and schemas, to represent resource descriptions and schemas under the 1. INTRODUCTION able information resources and the proliferation of descripable and machine understandable. Due to its exible model, HTML pages or XML data) to dieren t target communities [5], the Dublin Core [25], or the Web Service Description RDF is playing a central role in the next evolution step of vocabularies can be easily extended to meet the description from PDF or Word documents, e-mail or audio/video les to RDF, as well as, emerging application standards for Web tion services in various user communities, lead nowadays to graphs in which nodes are called resources (or literals) and of schemas [6] i.e., vocabularies of labels for graph nodes nels, etc.) about resources of quite diverse nature (ranging (i.e., classes) and edges (i.e., properties). Furthermore, these needs of specic (sub-)comm unities while preserving the aumation resources (e.g., sites, documents, data, images, etc.) of application contexts. To interpret resource descriptions syndication - and hence, automated processing - in a variety the provision of various kinds of metadata (for administrapressing metadata in a form that is both humanly readavailable on corporate intranets or the Internet. The ReLanguage [26]). In a nutshell, the growing number of availlarge volumes of RDF metadata (e.g., the Open Directory the Web - termed the Semantic Web. Indeed, RDF/S enable tonomy of description services for each (sub-)community. data and services syndication (e.g., the RDF Site Summary within or across communities, RDF allows for the denition and Web Portals (e.g., Open Directory, CNET, XMLTree1) Metadata are widely used in order to fully exploit infordata. More precisely, RDF provides i) a Standard Repre(corporate, inter-enterprise, e-marketplace, etc.). The most descriptions for the same Web resources, enabling content tion, recommendation, content rating, site maps, push chansentation Language for metadata based on directed labeled edges are called properties and ii) an XML syntax for exMany content providers (e.g., ABCNews, CNN, Time Inc.) ing the creation and exchange of metadata as any other Web distinctive RDF feature is its ability to suport superimposed source Description Framework (RDF) [17] aims at facilitator browsers (e.g., Netscape 6.0, W3C Amaya) already adopt

Permission to make digital or hard copies of all or part 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 bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission by the authors.

Semantic Web Workshop 2001 Hongkong, China Copyright by the authors. ture the semantics of web resources, while the second is inious information resources. Portals may be distinguished all cases, the key Portal component is the Knowledge Caterarchies of topics to semantically classify Web resources. we will hereforth omit the namespaces as well as the paths that ODP hierarchies are not deep (the maximum depth is enabling the development and maintenance of specic comtals, like Yahoo! or Open Directory (ODP), use huge hi13) while the number of direct subclasses of a class varies topics). The former, is represented using the RDF subclass query language can be used to provide declarative access to tions is determined by the corresponding namespace deninon-unique names) prexing class and propert y names. Figclassify and access, in a semantically meaningful way, varPortals are nowadays becoming increasingly popular by [12] on corporate intranets or the Internet. Such Commuaccording to the breadth of the target community (e.g., tardescribed resources. In Section 5 we will illustrate how our Specically , Table 1 lists statistics of 15 (out of 16) ODP tal Catalogs can be easily and eectiv elly represented using hierarchies (version of 16-01-2001), comprising 252840 toption of each schema, e.g., ns1 (www.dmoz.org/topics.rdfs) from the root of the topic hierarchies (since topics have tended for Portal administrators. The scope of the declaraure 2 depicts 2 (out of the 16) hierarchies in the ODP topics RDF/S. hierarchy (e.g., subtopics) or across hierarchies (e.g., related nity Web Portals essentially provide the means to select, For instance, the catalogs of Internet (or horizontal) Porics of the catalog to locate resources of interest, or issue a of resources are usually created using an OCLC Dublin-Core alog holding descriptive information, i.e., metadata, about relationships exist between these classes either within a topic The middle part of Figure 2 depicts the two schemas emaverage is around 4.02). Additionally, various administranamespaces in the upper part of Figure 2). Various semantic geting horizontal or vertical markets), and the complexity the catalog content. In the sequel, we illustrate how Porployed by the Open Directory: the rst is in tended to capand ns2 (www.oclc.com/dublincore.rdfs). For simplicity, like schema [25]. Users can either navigate through the topgreatly (the maximum number of subclasses is 314 but the of information resources (e.g., sites, documents, data). In represented as RDF classes (see the rdf and rdfs default the community resources. full-text query on topic names and the URIs or the titles of tive metadata (e.g., titles, mime-types, modication dates) schema, namely, Regional and Recreation, whose topics are ics under which 1770781 sites are classied. We can observe munities of interest (e.g., enterprise, professional, trading) Portal of Netscape exports in RDF around 170M of Subject storing and querying voluminous RDF descriptions, the Open Directory Portal exported in RDF (see Section 2). the manipulated desription bases and the underlying time other approaches using a monolithic table [20, In this paper we present ICS-FORTH RDFSuite [14], a suite under the form of triples.

Section 3 introduces a formal data model for descripto benet from three decades of researc h in database techresentation of several interconnected RDF schemas, as well as the introduction of a graph instantiation mechfor loading resource descriptions in an object-relational DBMS (ORDBMS) by exploiting the available RDF Section 6 illustrates the performance of RSSDB for performs both in storage volumes and query execution Finally, Section 7 presents our conclusions and draws directions for further research.

The paper makes the following contributions: descriptions at any time whilst it can be customized schemas with existing low-level APIs and le-based impleanism permitting multiple classication of resources.

Section 5 sketches the functionality of a declarative ing (RDF Query Language-RQL) using an object-relational performance of RDFSuite we use as testbed the catalog of 18, 7] to represent resource descriptions and schemas RDF application scenario. ical independence for RDF description bases. In this way, such as the ODP catalog. In particular, RSSDB outof tools for RDF validation (Validating RDF Parser-VRP), language, called RQL, for querying both RDF demanaging such voluminous RDF resource descriptions and tion bases created according to the RDF Model & SynSection 4 presents our persistent RDF Store (RSSDB) DBMS (see Figure 1). To illustrate the functionality and the as much as possible to the underlying ORDBMS. in several ways according to the specicities of both tion). The main challenge for this model is the repof determining how to eÆciently store or access them to the RDF applications have to specify in a high-level language maintenance of real-scale RDF applications. Still, we want storage (RDF Schema Specic DataBase -RSSDB), and queryibility of RDF in rening schemas and/or enriching schema knowledge. Our approach preserves the exmentations [22] does not ensure fast deployment and easy of RQL on top of RSSDB is described with emphasis tax and Schema specications [17, 6] (without reicanology to support declarative access and logical and physscriptions and related schemas. The implementation on how the RQL interpreter pushes query evaluation underlying RDF database engine.

Topics and 700M of indexed URIs). It becomes evident that only which resources need to be accessed, leaving the task 2. THE OPEN DIRECTORY CATALOG can be inherited (e.g., to subclasses of ExtResource), while prove storage volumes and optimize queries on these graphs. proposed models for semistructured or XML databases [1]: tional (e.g., the property le size is not used), can be multiresent superimposed descriptions of community resources, ily pairwise related by specialization: the instances of of XML trees (i.e., there is no root XML node). tion base through one or the union of all related schemas. It high-level queries on RDF/S graphs, while several physical Note also that due to multiple classication, resources may have quite irregular descriptions modeled only through Hence, RDF properties are unordered (e.g., the property title). Therefore existing proposals cannot be used, as an exception mechanism a la SGML [15]. Last but not such, to manipulate RDF description bases. In the sequel, Classes do not dene object or relation types: an ina class may have quite dieren t properties associated we will present a logical data model allowing users to issue least, semistructured or XML models can’t distinguish beunion of these properties is dened; representations can be used by the underlying DBMS to imResources may belong to dier ent classes not necessarstance of a class is just a resource URI; with them, while there is no other class on which the title can appear before or after the property description), opwhile preserving a conceptually unied view of the descripfrom standard object or relational models [3] or the recently set of constraints (domain and range compatibility). resources can be multiply classied (e.g., &r4). These modeling primitives provide all the exibilit y we need to repbecomes clear that the RDF data model diers substan tially valued (e.g., we have two description properties), and they tween entity (e.g., ExtResource) and property labels (e.g., Properties may also be rene d by respecting a minimal s n o i t p i r c s e d e c r u o s e R and two description properties with values \OÆcial site of (i.e., reic Finally, both RDF graph schemas and ation2).

RDF statements. Although not illustrated in Figure 2, RDF literal properties: a title property with value \Disneyland" (e.g., title), and an object (e.g., \Disneyland "). The subFinally, the ODP administrative metadata schema captures &r4). For instance, &r4 is a resource classied under both value (i.e., node) form a statement. Each statement is repdescriptions can be serialized in XML using various forests Using these schemas, we can see in the lower part of Fignode) together with a named property (i.e., edge) and its also supports structured values called containers (e.g., Bags) specialization) with the domain and range of the predicate class of Paris) and the latter using an RDF property named respectively. In the RDF jargon, a specic resource (i.e., resented by a triple having a subject (e.g., &r4), a predicate various descriptive elements of Dublin-Core [25] (with the related (e.g., between the classes Ile-de-France and Hotel). the classes Paris and ExtResource and has three associated ject and object should be of a class compatible (under class ure 2, the descriptions created for four sites (resources &r1(e.g., &r4 is of type ExtResource). In the rest of the pato represent attributes of resources as well as relationships between resources. Properties can also be organized in taxfor grouping statements, as well as, higher-order statements execption of the Subject element), as literal RDF properties dened on class ExtResource. Note that properties serve relationship (e.g., Travel is a, not necessarily direct, subper, the term description base will be used to denote a set of Disneyland Paris" and \Site oÆciel de Disneyland Paris" onomies in a manner similar to the organization of classes. 5.3 5.56 4.8 5.5 5 5.4 5.9 6.53 5.3 6.3 5 5.87 7.13 4.8 7.7 3.1 Additional RDF Constraints 3. A FORMAL DATA MODEL FOR RDF domain(p) 2 C and range(p) 2 C [ L. We denote by H = using class names or literal types so that: for each p 2 P , and domain(p2) range(p1) range(p2)3. and property names P P whose scope is determined by (N; ) a hierarchy of class and property names, where N = such that: if 2 P and then p1; p2 p1 p2, domain(p1) Each RDF schema uses a nite set of class names C C one or more namespaces. Property types are then dened C [ P . H is well-formed if is a smallest partial ordering 3.2 RDF Typing System relation is dened in RDF/S (e.g., bet ween bags of dieren t while properties as binary relations of type f[ U ; U ]g (for types). The set of all type names is denoted by T . meaningless). Furthermore, all types are mutually exclusive Sequence type, (:) is the Alternative type, and U is the type relationships) or f[ U ; (for attributes). Furthermore, L]g for resource U Alternatives capture the semantics of RIs4. 2 + : : : + Unlike traditional object data models, RDF n)]. (e.g., a literal value cannot also be a bag) and no subtyping The proposed type system oers all the arsenal w e need and alternatives, labels can be omitted (for bags, labels are tions (e.g., as a bag of sequences). Finally, assignment of a can be dened as heterogeneous of type [( 1 + sequences5 ples denoted by : : : ; (where is of some type [v1; v2; vn] vi i) union (or variant) types [8], and they are also ordered (i.e., classes and properties. For instance, unnamed ordered tuclasses are then represented as unary relations of type f U g RDF containers can also be used to represent n-ary relawhere L is a literal type in L, f:g is the Bag type, [:] is the to capture containers with both homogeneous and heterogeinteger labels play the role of union member markers). Since neous member types, as well as, to interpret RDF schema there exists a predened ordering of labels for sequences 3.3 RDF Description Bases and Schemas C \ P \ T = ; p 2 f1; 2; : : : g ) (v1) TBjSjA subPropertyOf relation is transitive: if p is subPropertyOf then the domain (range) of p is a subset of the domain (range) of p0 p0: if p belongs to the set f1; 2; : : : g then is an instance of either rdf:Bag or rdf:Seq or rdf:Alt: v1 p ) p p0; p0 p00 p00 p ) p 6= p0 p00 p P roperty p ) domain(p) ^ range(p) p0 domain(p0) range(p0) if v is a URI then it is an instance of one or more classes not related by : 8p; 2 P : p0; p00 A reied statemen t should have exactly one rdf:predicate, rdf:subject and rdf:object property Class is the root of the class hierarchy: Additionally, the direct extends of properties must be unique 8p 2 P; 2 [ p] : [v1; v2] subPropertyOf relation is antisymmetric: if p doesn’t belong to the set f1; 2; : : : g then belongs to the domain (range) of p: v1 (v2) subClassOf relation is transitive: if v is a literal value then it is an instance of one (and only one) Literal type: 8c; 2 C; v 2 ^[ c] ; c ) v 2= ^[ c0 c0 c0] 8c; 2 C : c0; c00 Domain and range of properties should be dened and they should be unique: Property is the root of the property hierarchy: c Class c ) c 6= c0 c0 8p 2 P; 2 C = domain(p)) ^ 2 C [ = range(p)) 9!c1 (c1 9!c2 TL (c2 8v 2 V : v 2 U ) (v) C if v is a container value then it is an instance of one (and only one) Container type: Literal, Resources and Container values are mutually exclusive c ) c c0; c0 c00 c00 v 2 L ) (v) and (v) is a singleton TL Class, Property and Type names are mutually exclusive v 2 V n (L [ U) ) (v) and (v) is a singleton TBjSjA p 2 P n f1; 2; : : : g ) 9n 2 n domain(p) ^ 9m 2 m range(p) (v1); (v2); subClassOf relation is antisymmetric: p 2 P n f1; 2; : : : g ) 2 P n f1; 2; : : : g; p ) 2= ^[ ) (8p0 p0 [v1; v2] p0] L \ U \ V n (L [ U) = ; 3.4 The Validating RDF Parser (VRP) 8www.w3.org/RDF/Implementations/SiRPAC 4. THE RDF SCHEMA SPECIFIC DATABASE rdf:Alt) and Literal types (e.g., string, integer, date). to allow the customization of the physical representation of The main novelty of our representation is the separation RDF metadata in the underlying ORDBMS. This is imporSubProperty tables, while the corresponding instance tables all class instances by a unique table Instances (see dashed Note that in the Triples table we distinguish between prop- traversals on subclass (or subproperty) hierarchies. both RDF schemas and resource descriptions using two ta- properties with range a class will remain unchanged. The scriptions, called SpecRepr, similar to the attribute-based ap- The benets of this SpecRepr variation are illustrated in Secthat the syntactic constraint (see Section 3) impos- RDF metadata in an ORDBMS, namely The 9Note PostgreSQL10. between schema labels is captured by the SubClass and experimented for the ODP catalog is the representation of This is not the case of other proposals employing a mono- sic SpecRepr could be the representation of properties with bles (see Figure 6-B), namely, Resources and Triples. The above representation is applicable to RDF schemas where bles store the URIs of resources while property tables store characteristics of employed schema classes and properties. and properties) by a unique id whereas the latter represents cialized. Finally, our SpecRepr opens the way for a more erties representing attributes (i.e., objvalue with literal val- 4.1 The RDF Description Loader sequel, it can be customized in several ways according to the also interesting to represent in a similar way all the instances the selection of specic tuples of the tables. Indices are also a limited number of tables. Furthermore, numerous tables Class), as well as, those of Container (e.g., rdf:Bag, rdf:Seq, Compared to GenRepr, our SpecRepr is exible enough the URIs of the source and target nodes of the property. Our main goal here is to reduce the total number of creproach proposed for storing XML data [13, 23]. SpecRepr tion 6 given that most ODP classes (i.e. topics) have few or bles Class and Property and on the attribute subid of the i.e. one for each topic) have a signican t overhead on the ORDBMS. Furthermore, a table Type is used to hold the with URI object values). Indices (i.e., B-trees) are conIt should be stressed that the taxonomic relationships its attributes, etc.). One of the possible variations we have relies on a schema specic representation of resource de- sources and the id’s of the classes in which resources belong. ported today by all In other words, RSSDB namely uri and classid, for keeping the uri’s of the re- ORDBMSs9. ing complete class and property denitions, ensures that loader has been implemented [4] in Java and communication specicities of both the manipulated description bases and of properties, but in general real-scale RDF schemas have names of RDF/S built-in types (e.g., rdf:Property, rdfs:- structed for all table attributes.

Figure 6-A), namely, Class, Property, SubClass and Subthrough specialization 10www.postgresql.org enriching descriptions at any time, and as we will see in the tain less than 30 URIs). For other RDF schemas it could be the statements made about the resources (including classes cleaver representation of class and property names using apthe RDF application scenario (i.e., querying functionality). more classes than properties. Another alternative to our bathe table hierarchy created in RSSDB can be only extended triples (captured by predid, subid and objid respectively). lationships of schema labels and enable to optimize recursive ues) and those relationships between resources (i.e., objid Figure 5 depicts the architecture of our system for loading instances of classes and properties. More precisely, class ta- resentation are required to take into account the specic Figure 4: The Validating RDF Parser (VRP) dened in an RDF sc hema. To save space when storing class constructed on the attributes lpart, nsid and id of the ta- (e.g., the ODP catalog implies the creation of 252840 tables, target of the above tables in order to speed up joins and commercial ORDBMSs (and not PostgreSQL) permit only of the RDF schema from data information, as well, as the tant since no representation is good for all purposes and in Indices are constructed on the attributes URI, source and ated instance tables. This is justied b y the fact that some and properties) under the form of predicate-subject-object propriate encoding systems that preserve the taxonomic represerves the exibilit y of RDF in rening sc hemas and/or no instances at all (more than 90% of the ODP topics conare also connected through the subtable relationship, sup- rectangular in Figure 6-A). This table has two attributes, resentation (i.e., for all RDF schemas), called GenRepr, of be added to the created class tables. The tables created for Property which capture the class and property hierarchies The core RDF/S model is represented by four tables (see distinction between unary and binary relations holding the most real-scale RDF applications variations of a basic reptables SubClass and SubProperty. response time of all queries (i.e., to nd and open a table, keeping the namespaces of the RDF Schemas stored in the lithic table [20, 18, 7] to represent RDF metadata under range a literal type, as attributes of the tables created for the and properties names, we also consider a table NameSpace Figure 5: RDF Schema Specic DataBase & Loader the form of triples. These approaches provide a generic rep- domain of this property. Consequently, new attributes will former represents each resource (including schema classes attribute-properties are single-valued and they are not spesource (variable Z of type resource URI) and target values topic names and other should be found in the descriptions The data path expression in the outer query iterates over the (variable T of type resource URI) of the title property. The where $X like "*Hotel*" ables prexed b y $ like $X) specializing the root Regional. value matches "*Opera*". Obviously, RQL schema queries data queries like Q, are not possible in current portals, since users. RQL aims to simplify such tasks, by providing powerThen, the ltering condition in the where clause will rethat similar information may be found under Recreation names of topics. Furthermore, compositions of schema and we can see in Figure 2, information about hotels in Paris of the resources. select Z portals may also be found under dieren t hierarchies, that The schema path expression in the from clause of the may be \connected" through related links. For example, as \." implies an implicit join condition between the extent of and $X < Paris)fYg.fZgtitlefTg hotels in Paris whose title matches "*Opera*". from Regionalf:$Xg ries) of a given topic, browsing should be continued by the class name and ranges over the result of the nested query. nested query, states that we are interested in classes (variQ: Find the resources under the hierarchy Regional, about such links are not necessarily bi-directional; thus, a user tain only the classes whose name matches "*Hotel*" and may also be found in the Recreation hierarchy. Moreover, starting from the Regional hierarchy may never nd out from (select $X are far more powerful than the corresponding topic queries schema knowledge is that the subtopics of Regional contain ful path expressions permitting smooth ltering/na vigation geographical information and a topic Paris is somewhere in To make things more complex, relevant information in nally, the result of Q will be a bag of resources whose title they are also subclasses of Paris (e.g., Hotel and Hotel Directories. Here, to get all relevant topics, the only required each class valuating Y and the resources valuating Z. Fiwhere T like "*Opera*" ed under the topic of in terest. Note that, in order to locate one cannot specify that some query terms should match the previous query can be expressed as follows: resources classied under the subtopics (e.g., Hotel Directoon both Portal schemas and resource descriptions. Then the of common portals, which allow only full-text queries on the the hierarchy. Thanks to RQL typing, variable Y is of type source-values of title and the proper instances of the class is rather straightforward: the class variable $X over all the subclasses of Regional and its values are ltered according case of the query Q presented previously. sions is left to PostgreSQL while the evaluation of RQL which the evaluation of the right expression depends on the to the conditions in the where clause. The operator Map on loop evaluation with values of variable Y passed from the sions interleave schema with data querying [9]. This is the (c) the evaluation engine, accessing RDF descriptions from top is a simple variable renaming for the iterator (Y ) dened evaluation to the underlying SQL3 engine. Then pushing pression (expressed in an object algebra a la [10]) to a single extents (W ) returned by the nested query, as shown in the (a) the parser, analyzing the syntax of queries; (b) the graph right branch. The connection between the two expressions lies on a full-edged ORDBMS like PostgreSQL, the goal Figure 7 (a) illustrates the algebraic translation of Q. The in the from clause is translated into a semi-join between the functions for traversing class and property hierarchies relies translation of the nested query - given in the left branch over the nested query result. Then, the data path expression The RQL interpreter, implemented in C++, consists of main diÆculty in translating an entire RQL algebraic exof the RQL optimizer is to push as much as possible query typing and interdependencies of involved expressions; and on the existence of appropriate indices (see Section 6). The constructor, capturing the semantics of queries in terms of the underlying database [16]. Since our implementation reevaluation of the left one). Djoin corresponds to a nested selections or reordering joins to evaluate RQL path expresis captured by a Djoin operation (i.e., a dependent join in SQL3 query is due to the fact that most RQL path expresdown to PostgreSQL, leaving the evaluation of the Djoin to select Z.source procedural SQL functions) as follows: scan operation on class instances (^(Y )[W ]) with a selection on the Instances relation, as shown in Figure 7 (b). Then, is expressible in SQL3 (subclassof and issubclassof are shown in Figure 7 (c). As one can see, the whole query Note that the * indicates an extended interpretation of tables, according to the subtable hierarchy. Thus, the whole right into a join and pushing the selection criterium C = mizations, as pushing down selections, join reordering etc.

Y up to the Djoin operation leads to the evaluation plan the interpreter. However, transforming the semijoin on the query can be pushed down to PostgreSQL, and the Postfrom subclassesof(Regional) Y, Intances I, title* Z each of the subqueries in the shadow boxes can be pushed left-hand side (i.e., nested query) to the right-hand side. where issubclassOf(X, Paris) and X like \*Hotel" and Z.target like \*Opera*" I.classid=Y and Z.source=I.uri and greSQL optimizer can perform on further (common) optiUsing the database representation, we can substitute the 5.1 RQL Interpreter 6. PERFORMANCE TESTS respectively. The (semi-)joins involved in the evaluation of three dieren t result cases per query. The main observaa matter of fact, these query templates depict the core funcexecution time of a query. Table 4 illustrates the obtained desired querying functionality for RDF description bases, queries for specic resource descriptions (e.g., Q10, Q11). As large Triples table are required. namely: a) pure schema queries on class and property de(0.124%) for table Triples (SubClass) in the three cases To get signican t experimental results, we carried out all outperforming GenRepr by a factor of up to 3.73 approxthe table names of Figure 6). This benchmark illustrates the apparent in the cases where self-join computations on the imately. This is clearly illustrated in Q1, where one tu5%) and Property (selectivity 20%) using index and seclasses which represent in GenRepr (SpecRepr) selectivities the database buers and then nine times to get the a verage tionality of RQL presented in the previous section. in queries Q1, Q2, Q5, Q7, Q10 and Q11, with SpecRepr pressions in the two representations (capital letters abreviate GenRepr and SpecRepr exhibit comparable performance queries considered. The deviation in performance is more that the time required to lter a table in both represen taof 1.7e-5% (3.955e-4%), 5.14e-4% (1,19e-2%) and 5.38e-3% queries Q5, Q7 and Q11 incur an additional cost for GenRepr, we have experimented with classes having 1, 30, 314 subple is selected from both table Triples (selectivity 1,7etion is that SpecRepr outperforms GenRepr for all types of tions depends on the number of tuples in the query results: ing available schema knowledge (e.g., Q5-Q9); and c) schema benchmark queries several times: one initially to warm up nitions (e.g., Q1-Q4); b) queries on resource descriptions usTable 3 describes the RDF query templates that we used quential scans respectively. As expected, in Q2, we can see execution time (in sec) for both representations in up to for our experiments, as well as their respective algebraic exquired in SpecRepr for storing the URIs of resources in the Triples (i.e., 1770781 extra tuples). Although not illusconstructed, the average storage volumes per data triple bethe ODP schema is 1539 sec. store a data triple is 0.123KB in both representations. This obtain better storage volumes in the SpecRepr by encoding the number of data triples. The average space required to respectively. Despite the use of ids, the indexes in GenRepr triple requires on average 0.086KB versus 0.1582KB in the 44MB. The size of the Instances table is 150MB (1770781 (GenRepr) and the average loading times become 0.0025 sec property tables compensates for the extra tuples stored for come 0.2566KB (SpecRepr) and 0.2706KB (GenRepr) while Catalog). We can observe that in both representations the size is 32MB for Class (252825 tuples), 8KB for Property that for each class denition in GenRepr three tuples are 6.2 Querying Figure 8: Statistics for Loading Schema and Resource Descriptions The database size required for the resource descriptions GenRepr. This is mainly attributed to the space saved in tuples) whereas the size of indexes created on it is 140 MB. larger in GenRepr because of extra (252825) tuples stored. each resource description in GenRepr. Note that we could the average loading time become 0.0043 sec and 0.00457 sec loading a data triple is about 0.0033 sec and 0.0039 sec the URIs. The tests also show that the average time for To summarize, after loading the entire ODP catalog, the triples. The tests show that, in SpecRepr, each schema SubProperty, and the total size of indexes on these tables is but this solution comes with extra loading and join costs stored: one in table Resources and two tuples in table take up more space because of: a) the extra tuples stored size of tables in GenRepr is 545MB for Triples (5835756 tutriples: one for the class itself and one for its unique super- ples), 202MB for Resources (2022869 tuples) and the total (between the class and property tables) for the retrieval of and 0.00317 sec respectively. The total validation time for the DBMS size in both representations scales linearly with the resource URIs as integers, as we did in the GenRepr, When indices are constructed, the average storage volumes class (multiple class specialization is not used in the ODP size of indexes on these tables is 900MB. In SpecRepr, the b) the index constructed on the predid attribute for which per schema triple become 0.1734KB (SpecRepr) and 0.3062KB trated here, the average time for loading a schema triple is SpecRepr due to the Namespace table, as well as to the fact is shown in the rightmost graph of Figure 8. As expected, about 0.0021 sec and 0.0025 sec respectively in the two repthere is no corresponding index in SpecRepr. size of the DBMS scales linearly with the number of schema (5 tuples), 11MB for SubClass (252825 tuples) and 0MB for respectively in the two representations. When indices are should not appear as a surprise since the extra space reresentations. The time required to store a schema triple is ) predid=6^objid=clsid(T