ACM SIGSOFT Software Engineering Notes Page 1 May 2006 Volume 31 Number 3

Transcription

1 ACM SIGSOFT Software Engneerng Notes Page 1 May 2006 Volume 31 Number 3 Abstract STCIM: A Dynamc Granularty Orented and Stablty Based Component Identfcaton Method Zhong-Je Wang, De-Chen Zhan, Xao-Fe Xu School of Computer Scence and Technology, Harbn Insttute of Technology, Harbn, Chna {rany, dechen, xaofe@ht.edu.cn Among recent developments n the feld of software reuse has been the ncreasng reuse of coarse-graned components, and t has been proved that granularty has great mpact on component s reuse performance. However, prevous studes have gnored rgorous and effectve methods to support coarse-graned component dentfcaton and desgn, partcularly granularty optmzaton desgn. In ths paper, a stablty-based component dentfcaton method, STCIM, s presented to resolve ths problem. Frst a feature-orented component model and the correspondng component granularty metrcs are brefly presented. By establshng mappngs between busness model space and component space, component desgn process may be regarded as the process of decomposton, abstracton and composton of busness model elements, wth four dfferent mappng strateges dscussed to obtan dynamc component granulartes. Furthermore, t s thought that component granularty s closely correlatve to the stablty of busness models: the more stable the busness model, the larger the correspondng component granularty may be. A metrcs for model stablty wth three factors,.e., number of somers, stablty entropy and somer smlarty, s presented, and the correspondng component dentfcaton algorthm based on Most Stable Set s dscussed n detals. Fnally a practcal case s descrbed to valdate the method n ths paper. Keywords: busness component, component dentfcaton, feature modelng, stablty, dynamc granularty 1 Introducton Wth the rapd development of software reuse theores and technques, the evoluton of reusable software artfacts have ranged from functons, objects and classes n object-orented programmng, to components, frameworks, desgn patterns [1][2], servces [3] and software archtectures [4] n today s Internet-based envronments. The granularty of reusable artfacts s contnuously ncreasng [5]. Currently, coarse-graned reuse, such as component-based development [5][6] and servce-orented computng [7], has been the prmary drectons n software reuse [2]. Granularty depcts the scale and complexty of a reusable artfact and s dffcult to be precsely measured. Qualtatvely speakng, components may be classfed nto fne-graned, medum-graned and coarse-graned ones [8], called granularty level. For example, a process component provdes a sngle, dscrete busness process, whle an actvty component provdes one or more busness actvtes, therefore the granularty of the former s hgher than the latter. Dfferent components n the same granularty level stll have dfferent granulartes, whch can be evaluated n two ways: the number of functons that a component can provde outward by ts nterfaces, and the number of reusable enttes contaned n the component. To some up, the coarser granularty a component has, the more functons t may provde, or the more enttes t has mplemented. The reason that granularty s consdered a key concept n software reuse s that t nfluences components reuse performance to a great extent [1][6]. Components wth fner granularty have wder range of reuse and hgher reusablty, however, to realze a large and complex system, numerous fne-graned components have to be tme-consumngly composed together, therefore the reuse effcency s lower and reuse cost s neffcent. In comparson, coarser granularty components certanly have hgher reuse effcency and lower reuse cost. On the other hand, fne-graned components, such as busness object components, depct those stable and core busness elements n an enterprse and have experenced few changes durng ther whole lfecycle, therefore, when these components are reused, t s not requred to be modfed or reconfgured on a large scale. However, coarse-graned components, e.g., busness actvty components, busness process components, contan a mass of unstable busness elements such as busness rules [9], whch usually need to be changed regularly wth rapdly changng busness requrements. Therefore, when these components are reused, they have to be reconfgured or modfed to a great extent, and wll lead to hgher reuse cost. Mnmzng unstable elements contaned n a component and maxmzng component granularty wll enable us to explot ts advantages and enhance ts reusablty at the same tme. Although coarse-graned reuse s the prevalng trend, coarser granularty does not necessarly equal better qualty. Therefore, the proper specfcaton of components to acheve maxmum enhancement of ther granularty whle mnmzng the lmtatons nherent n coarse-graned components s a crtcal challenge n the component dentfcaton and desgn process, whch ths paper s gong to address. Component dentfcaton s a prmary msson n the phase of doman engneerng, n whch, doman busness models are decomposed and clustered nto a set of reusable components [10][11]. By reusng these components n the phase of applcaton engneerng, software systems that support specfc busness models can be mplemented easly. Therefore, there s a strct and b-drectonal mappng between busness models and components,.e., components are the software representaton of busness models, and busness models are the semantcs representaton of component models. Due to the exstence of ths mappng, component granularty s closely related to the propertes of busness models. Ths knd of mappngs should satsfy semantc consstency. When components are reused to construct specfc busness models (or, software systems), f ther semantcs cannot reach consstency,.e., the functons that a component provdes cannot satsfy (or conflcts wth) the requrements of the models, t becomes necessary to adjust nner structure and semantcs of these components, thus

2 ACM SIGSOFT Software Engneerng Notes Page 2 May 2006 Volume 31 Number 3 causng hgher reuse cost. From the statstcal vew, f a busness model n a specfc busness doman needs frequent changes, then, those components obtaned from ths busness model are sure to have a hgh probablty of frequent changes, too. We refer to the probablty that a busness model or a component remans stable n a gven perod of tme as stablty. Component stablty and busness model stablty are closely related. The hgher stablty a busness model has the more unlkely those components obtaned from the busness model wll change frequently, and vce versa. Therefore, t s better that a component formed wth hghly stable busness elements results n a coarser granularty, whle frequent changng busness elements wll result n fner-graned components. Based on ths dea, a component dentfcaton method, STCIM, s presented n ths paper for granularty optmzaton based on busness model stablty. In ths method, the stablty of doman busness models s used as the decson nformaton n the process of component dentfcaton and granularty optmzaton. By calculatng the stablty of every busness element and the degree of stablty dependency between these elements, we map those stable elements nto coarser-graned components, and for those nstable elements, we teratvely decompose them nto several stable clusters and map every one of them to a fner-graned component. The method could fnally brng about mult-granularty component coexstence. The rest of ths paper s organzed as follows. In secton 2 related works n lteratures are brefly ntroduced. In secton 3, we present a feature-orented busness model and the correspondng defnton of granularty based on feature space. In secton 4, we dscuss the mappng between busness model space and component space, wth four mappng strateges emphaszed. In secton 5 the defnton of stablty and stablty dependency, and the correspondng metrcs, are descrbed n detal. A component desgn method based on stablty, and a practcal case, s fnally presented n secton 6. The concluson s gven n secton 7. 2 Related works 2.1 Component Granularty Researchers have proposed varous methods to defne and measure component granularty qualtatvely and quanttatvely. Generally speakng, component granularty s a metrcs that s used to descrbe the degree of coarseness or thnness of those functons provded by a component [12]. The smplest evaluaton way of component granularty s the length of nteror busness logc of the component, e.g., number of program lnes [13]. In fact component granularty s dffcult to be precsely measured. If a component s not self-contaned,.e., n order to use the component, some other components must also be reused at the same tme, then t s a fne-graned component, otherwse t s a coarse-graned component [14]. In an approxmate way, an entty component s granularty can be calculated by the number of busness enttes contaned n ths component. For a process component, t s granularty s ncreasng lnearly wth the number of branches and actvtes n the process dagram realzed by the component [15], and can be calculated by the formula Granularty(PC)= u PC [α Branch(u) 2 +β Actvty(u)]. Dfferent types of components are n dfferent granularty levels, and a coarse-graned component can be mplemented by the composton of a set of fne-graned components. Therefore, granularty s a concept that should be recursve defned. Generally speakng, component granularty can be calculated by the synthess of granularty of sub-elements contaned n t. 2.2 Doman engneerng and component dentfcaton method Component dentfcaton and desgn s a prmary msson n doman engneerng (DE). Component desgners carry out doman analyss from a set of smlar requrements n a specfc doman to fnd out the commonaltes and varabltes n ths doman and construct DSSA [16][17]. Accordng to these results, reusable busness semantcs are obtaned and the correspondng component specfcatons can be desgned [2]. Some typcal and wdely appled DEmethods nclude FODA [18], FORM [19], ODM [20], RSEB [21], etc. Actually the basc dea of DE s consstent wth Model-Drven Archtecture (MDA) [22]. Doman models depcts the common busness requrements n a doman, and they belong to Computaton-ndependent Model (CIM), whle component-based software models may be consdered as a knd of Platform-Independent Model (PIM), therefore, component dentfcaton can be regarded as the transformaton from CIM to PIM n MDA. Research concernng ths aspect s not extremely mature and complete. It s not completely at wll to cluster busness models nto compo- nterest wth the deep research on Componentnents, and certan prncples must be followed. The majorty of current component desgn prncples mostly come from desgn prncples of class and class package n object-orented methods, such as Open-Close Prncple (OCP), Common Reuse Prncple (CRP), Acyclc Dependency Prncple (ADP), Sngle Responsblty Prncple (SRP), etc [23]. But, a component s not a smple aggregaton of classes; therefore these prncples are not fully sutable for component desgn. In addton, these prncples can be prmarly used n the desgn of entty components, and s dffcult to nstruct the dentfcaton of process components. On the other hand, these prncples are unable to accurately answer the queston that what knd of granularty can lead to best reuse performance, and there also lack of strct and effectve methods to support the dentfcaton and desgn of coarse-graned components. The problem of component dentfcaton and desgn has aroused the wdespread based Software Engneerng (CBSE) [2] n md 1990s. Intally, component dentfcaton was regarded as a phase n doman engneerng. These methods, e.g., FODA [18] /FORM [19], emphasze partcularly on the reusablty of DSSA and components compatblty and adaptablty between smlar applcaton famles n a doman, but do not pay much attenton to the optmzaton on other factors, such as reuse cost, reuse effcency, etc, and granularty s not consdered, nether. In later researches, t has obtaned full attentons as an ndependent problem. Startng from consderng component reuse cost, re- searchers try to decompose busness models nto a set of submodels accordng to those prncples such as hgh-coheson and low-couplng [6] by calculatng the overall relevance between busness elements, and encapsulate every sub-model nto one

3 ACM SIGSOFT Software Engneerng Notes Page 3 May 2006 Volume 31 Number 3 component. Typcal representatves nclude CRWD matrx based COMO [24] and O2BC [25] methods, and many clusterng analyss based methods, e.g., mnmal spannng tree technques, graph clusterng, and heurstc algorthms, etc, whch are mported from other felds such as Artfcal Intellgence and Reverse Engneerng. Components obtaned by ths knd of methods have loosely coupled semantc structure, whch ensures low reuse cost, however, n these methods, component reusablty and the ablty to adapt to changes are unable to guarantee to keep n a hgh level, and, optmzaton on granularty bascally does not nvolve, nether. Actually, the obtaned components are all fxed n one or several granularty levels, and component desgners even try to seek the balance between granulartes of dfferent components by ntenton [15],.e., f a component has coarser granularty than others, t wll be decomposed to decrease the granularty, whch s wdely dvergent wth current tendency, or conventons, of coarse-graned component reuse and runs n an opposte drecton. For the dentfcaton of coarse-graned entty components, we can compose multple fne-graned entty components by nhertance and composton relatonshps between them to get coarse-graned ones [26], but for process components, there are no good methods at present yet. 2.3 Stablty Stablty s orgnally a concept n Physcs, and was defned as the degree of the qualty or attrbute of beng frm, steadfast and free from change or varaton when outsde condtons changes [27]. Afterwards, t was ntroduced to the doman of System Theory, and was used to descrbe the ablty of a system, when kept under specfed condtons, to mantan a stated property value wthn specfed lmts for a specfed perod of tme [27]. For an enterprse, ts management patterns may be expressed by a set of busness models. And along wth the changes on nternal requrements and external envronments, busness models may also change [28]. Dfferent models have dfferent stablty. Some busness elements seldom change, so they have hgher stablty, whle other elements wll frequent change, so they have lower stablty. For example, the organzaton structures, poston settngs, executon rules n busness processes, are all easly changed [29]. Stablty could be measured as a volume of changes n the model needed to reflect changes n the realty. In the methodology of Sx Sgma, the stablty of a busness process model s defned as the ablty of the process to perform n a predctable manner over tme, and s evaluated by the tool Run Chart [30]. In the feld of software engneerng, stablty s used to descrbe the ablty of a software artfact to keep unchanged along wth the tme, or the ablty to adapt to changes by ts flexble confguraton mechansm [31][32]. In [33] a stablty metrcs for software archtecture and components based on Case-based Reasonng (CBR) s presented, n whch, the stablty s measured by the data produced n the process of software evoluton. Researchers have already obtaned the concluson that those components that are more frequently reused are more stable than the ones that are seldom reused [34]. So far, a majorty of doman analyss methods do not nvolve the factor of stablty n component desgn. Ths s because that there exsts a basc hypothess that all the commonaltes n doman busness models are all stable along wth tme [35]. But actually, ths hypothess s not true because any software artfacts wll change more or less wth tme. Hamza and Fayad dd some deep research on the stablty of software systems, and they have proposed a Stablty-Orented Doman Analyss (SODA) method [32][35], n whch the commonaltes are classfed nto endurng ones and nstable ones, accordngly, the software system s separated nto three layers: core and essental Endurng Busness Themes (EBTs) seldom changed Busness Objects (BOs) and frequently changed Industral Objects (IOs). In ths way, the fnal components have clear stablty dfference. But SODA does not suggest constructve strateges on component granularty optmzaton, nether. 3 Feature-orented component model and granularty metrcs Feature modelng s the actvty of modelng the common and the varable propertes of concepts and ther nterdependences and organzng them nto a coherent model referred to as a feature model. Feature-orented method has been wdely appled n the feld of software reuse long tme before, e.g., FODA[18], FORM [19], etc. Feature, as a tool, s used to depct the commonaltes and varabltes among related busness n a gven busness doman to form doman feature model for reuse. Reusable busness models contan nherently more varablty than concrete models and feature modelng s the key technque for dentfyng and capturng varablty. Because there exst an ntrnsc consstency between doman features and the servces provded by components, feature modelng method s also used to descrbe component models [12], and makes component models and doman models n a unform semantcs space, thus creates a drect mappng between them two. In addton, some of characterstcs of feature, such as herarchy, expansblty and mult-dmenson, make t more sutable to express the varatons n doman models and component models than other technologes. In ths secton, a feature-orented component model s presented. It s a sem-semantc component specfcaton model, wth emphass on the representaton of the commonaltes and varabltes of busness semantcs, and dependences between them. We try to use feature space as a tool to create a drect mappng between component and busness models,.e., busness models and component models are unformly expressed as feature space, and a component s regarded as a sub-space of busness model s feature space. Frst some taxonomy of feature and feature dependency are ntroduced, and then the feature-orented busness model s presented. Fnally some metrcs for measure component granularty, and relatonshps between component granularty and component reuse performance are dscussed brefly. 3.1 Feature and feature dependency Here we brefly ntroduce the concepts of feature and feature space n tradtonal feature modelng methods. Defnton 3.1. (feature[12]). Feature s an ontology that s used to

4 ACM SIGSOFT Software Engneerng Notes Page 4 May 2006 Volume 31 Number 3 descrbe the knowledge of external world, and s represented as Terms or Concepts descrbng the servces suppled by a specfc busness doman. Defnton 3.2. (feature space[12]). Feature space s a semantcs space composed wth a set of related features n a specfc doman, and dependences between these features, denoted as Ω=<F, D>, where F s set of features, and D s set of feature dependences between features n F. Ω s usually represented as the form of a feature tree, n whch, there exsts one and only one root node f root as the root feature of Ω, and the conjont two features n ths tree are respectvely called parent feature and chld feature. Each chld feature specfes a porton of ts parent feature s semantcs. The relatonshp between father feature and chld feature s a knd of aggregaton that s used to descrbe the whole-part assocaton (WPA) between features, and makes the feature space form a herarchcal structure. Denote chld(f), parent(f), ancestor(f), descendant(f) and sblng(f) as f s chld feature set, parent feature set, ancestor feature set, descendant feature set and sblng feature set, respectvely, and we have U ( g) ancestor, f f root ancestor( f ) = g parent( f ) Φ, f = f root U descendant( g), chld( f ) Φ descendant( f ) = g chld ( f ) Φ, chld f = Φ Defnton 3.8. (FSD). FSD refers to the semantcs relatonshps between features. Accordng to the types of features, FSD possbly have multple types, e.g., the temporal constrants between busness operaton features, the assocaton, nhertance and compos- If a feature f has multple parent features, then f s a common feature. We can construct a copy of f and all ts decedent features for f s every parent feature so as to ensure the feature tree satsfy the characterstcs of a common tree. Defnton 3.3. (feature tem). A feature Item s an nstance of a specfc feature, descrbes the feature s one possble value under a gven busness envronment, and reflects the varablty of the feature. Let dom(f) denote the set of all feature tems of feature f, and s called the doman of f. For τ dom(f), τ s called a value of f. A feature s an abstracton of all ts feature tems, and there exsts a generalzaton-nstantaton assocaton (GIA) between a feature and ts tems. Defnton 3.4. (feature nstantaton). Feature f s nstantaton s the process of selectng one feature tem from ts doman dom(f) to satsfy a requrement context R, denoted as τ R (f). If the meanng of R s clear, τ(f) wll be used nstead of τ R (f). An nstance of feature f can be denoted by the nstances of ts chld features. For τ dom(f), there must Y chld(f),y={f 1, f 2,, f n, and for f Y, there must at least exst one feature tem τ dom(f ) that makes τ s determned by τ 1,τ 2,,τ n unquely. Y s called the essental sub-feature set of τ, denoted as es_set(τ). Now we extend the tradtonal feature theory by proposng the defnton of feature dependency. A feature dependency (FD) descrbes the relatonshps between two related features n a feature space. Accordng to the structural and semantc relatonshps between two features, FD can be classfed nto four types: Feature Integrty Dependency (FID) Feature Value Dependency (FVD) Feature Mult-Value Dependency (FMVD) ( ) Feature Semantcs Dependency (FSD) Defnton 3.5. (FID). There exsts FID between feature f and feae set Y, f and only f Y chld(f), and for every feature tem τ of tur f, τ unquely determnes a set Y Y that satsfes Y es_set(τ), denoted as f Y. f Y can be refned nto four types, denoted as f M g, f O g, f S Y, and f T Y, respectvely. Mandatory FID: f M g τ dom(f), there must g es_set(τ); Optonal FID: f O g P,Q dom(f),p Q=dom(f),P Q=, for τ P, there exsts g es_set(τ), and for τ Q, there exsts g es_set(τ); Sngle selecton FID: f S Y Y dom(f) ; There exsts an clusterng {P 1, P 2,, P n of dom(f), n= Y, P P j =, =1 n P =dom(f), and for P, there exsts one and only one f Y, that makes τ P, f es_set(τ), and for f Y\{f, f es_set(τ); T Multple selecton FID: f Y There exsts an clusterng {P 1,P 2,,P n of dom(f), whch makes P P j=, =1 n P =dom(f), and for P, Y Y, for τ P, f j Y, there exsts f j es_set(τ), and for f j Y\Y, f j es_set(τ). FI D can be regarded as the dependences between the Values of pare nt feature and the Type of ts chld features, and s called Value-Type dependency,.e., one feature tem of parent feature determnes whch of sub-features are the essental parts of the parfeature. FID depcts the structural ntegrty relatonshps be- ent tween parent and chld features by the choce of whether a chld feature should be contaned n the parent feature under dfferent crcumstances. It s also the only feature dependency presented n tradtonal feature modelng methods. FID s knd of explct dependency, and can be expressed by the structure of feature space. Defnton 3.6. (FVD) Suppose the feature set X and Y are two subsets of F n feature space Ω, and for arbtrary two nstances t1, t2 n Ω s nstance set T(Ω), f t 1 [X]=t 2 [X] comes true, then t 1 [Y]=t 2 [Y] s also true, then Y s feature value dependent on X, or X feature determnes Y, denoted as X Y,.e., one nstance of X unquely determnes one nstance of Y. Defnton 3.7. (FMVD). Suppose the feature set X, Y and Z are three subsets of F n feature space Ω, and Z=F X Y. The FMVD X Y comes true f and only f for every dentcal nstance of (X, Z) determnes a set of Y s nstance, and these nstances only le on the nstances of X, and does not relate to Z. FVD and FMVD depct the restrctons that must be satsfed when dfferent features are nstantated. The essence of them are same, they both express the dependences between feature tems of two set of features, and can be called Value-Value dependency,.e., the nstances of one feature set unquely of multply determne the nstances of another feature set. They generally appear be- tween sblng features.

5 ACM SIGSOFT Software Engneerng Notes Page 5 May 2006 Volume 31 Number 3 ton constrants between busness object features, the Event- Condton-Acton (ECA) constrants between busness actvty features. Generally, we use a predcaton P(X) to denote that features n X should satsfy P. The concrete expressons of P are determned consderng the concrete constrant types. FSD has no relatonshps wth feature nstantaton, and t depcts the semantcs constrants between features. By the way, FVD, FMVD and FSD are mplct dependences. 3.2 Feature-orented component model A busness component defnes a sub space of a specfc busness doman. The feature-orented component model can be denoted as C<cd, f root, F, D, PS, RS>, n whch, cd s a unque dentfcaton of C; f root s the root feature, F s the feature set, f F, dom(f) 1, and D s the set of feature dependences between features n F; PS s the set of features that C provdes, RS s the set of features that C requres, and PS, RS F. Features n PS consttute the PROVIDE nterface of C, and features n RS consttute the REQUIRE nterface of C. For a normalzed component, t s often true that PS={f root,.e., C only provdes ts root feature, whch has the coarsest granularty, to ot her components, and f C need to provde other features n F, these features had better be separated out and form new ndepend- ent components. Component reusablty can be represented through the followng two aspects accordng to four types of feature dependences mentoned n secton 3.1: (1) Feature s varablty: Only n some certan stuatons, one feature s an essental consttuent of component feature space, and n other stuatons t s not. Accordng to FID, features can be dvded nto four types,.e., mandatory features, optonal features, sngleselecton features and multple-selecton features; (2) Feature tem s varablty: One feature can be nstantated as an arbtrary feature tem contaned n the feature s doman accordng to FVD/FMVD. The exstence of FVD/FMVD,.e., X Y or X Y, makes the nstantaton of X and Y cannot keep ndependent on each other, and t s necessary to nstantate features n Y accordng to the results of X s nstantaton to ensure the valdty of X Y (X Y). In fact, feature s varablty can be transformed to feature value s varablty n a smple way. For f whch s a non-mandatory feature, by addng one null feature tem I nto dom(f), f can now be regarded as a mandatory feature approxmately. If f should not be chosen n some crcumstances, f can be consdered to be nstantated as I. Under the premse of all the feature dependences n a component C, by choosng one specfc feature tem for every feature n C, C s nstantated and a set of component nstances are obtaned. Denote nstance(c) as the set of all the component nstances, and for f F, t nstance(c), denote ρ(f, t) as the correspondng feature tems of f n t. Component nstantaton s carred out when the component s practcally reused. In Fg.1 we present an example component that contans two fxed features f 1, f 2, three varable features f 3, f 4, f 5, and a root feature f root. There exst feature dependences {f 3 {f 4,{f 3 {f 5 between f 3, f 4, f 5, and dom(f 1 )={τ 31,τ 32,I, dom(f 2 )={τ 41,τ 41, dom(f 3 )={τ 51,I. Every feature tem of f 3, f 4, f 5 has a specfc mplementaton or a null mplementaton. C mpl(f1 ) mpl(f 2 ) PS f root VAR_PART(C) FIX_PART(C) {f 3 {f 5 { f 3 {f 4 f 1 f 2 f 3 f 4 τ 31 τ 32 I mpl(f 3,τ 32 ) mpl(f 3,τ 51 ) RS τ 41 τ 42 I τ 51 Legend fxed feature varable feature mplementaton of feature tem Fg.1 Feature-based component model and ts varaton mechansm Defnton 3.9. (feature s granularty level [12]). In the feature space of a specfc doman, we use the dfference between the layer of feature f and the root feature as f s granularty level, denoted as GL(f)=layer(f) layer(root). We have layer(root)=0, GL(root)=1, and layer(f)=layer(parent(f))+1. Features n the same granularty level usually have the same semantcs type, e.g., bus- ness actvtes, busness operatons, etc. 3.3 Component granularty Accordng to the general component model n last secton, here we propose some defntons of component granularty and ts metrcs, n whch feature s granularty s used to descrbe component granularty. Defnton (feature s granularty) A feature f s granularty G(f) s defned as the sum of all ts chld features granulartes,.e., G ( f ) ( f ) G = f chld( f ) 1,, f 5 chld chld ( f ) ( f ) Φ = Φ Because a component c can be regarded as a sub-space of doman feature space, c s granularty can be defned by the granulartes of features contaned n t. In secton 1 we have mentoned that the granulartes of components n the same granularty level can be measured from the followng two vews: the number of functons that a component can provde outsde by ts nterfaces, the number of sub enttes contaned n the component. They are called nterface granularty and mplementaton granularty, respectvely. From another vew, component granularty can also be classfed nto absolute granularty and relatve granularty. So we have four types of granularty, the defntons of whch are as follows: Interface Relatve Granularty: the number of features contaned n PS(c),.e., IRG () c = PS() c Interface Absolute Granularty: the mean value of granularty of all features n PS(c),.e., 1 IAG() c = ( ) () G f PS c f PS( c)

6 ACM SIGSOFT Software Engneerng Notes Page 6 May 2006 Volume 31 Number 3 Implementaton Relatve Granularty: the number of leaf fea- tures n components,.e., NRG c = f f F c, chld f = Φ () { () ( ) Implementaton Absolute Granularty: the granularty of f root,.e., NAG () c = G( f root ) Relatve granularty depcts the number of features a component provdes or contans, but t doesn t depct the dfference between these features granularty. Absolute granularty consders ths dfference. For a generc component, nterface granularty and mplementat on granularty s normally unform,.e., a component usually provdes the feature wth the coarsest granularty (root feature). If extra remarks are not added, the taxonomy component granularty referred n ths paper means mplementaton absolute granularty NAG(c). 3.4 Relatonshps between granularty and component reusablty Granularty has great nfluence to component reusablty, whch s reflected n three aspects,.e., reuse effcency, composton cost and change cost. Frstly the defntons of the three propertes are put forward. Defnton (Reuse effcency). Component c s reuse eff- of a cency REF(c) s defned as c s contrbuton to the assembly doman feature space cooperatng wth other components, and can measured by the proporton of c s feature space s sze n the total doman feature space, denoted as () REF c = F() c F ( Ω) It s easy to know that, the coarser c s granularty s, the more contrbuton c can make to the whole doman feature space, and therefore the hgher reuse effcency c has. Defnton (Composton cost) Component c s composton cost CPC(c) s defned as the cost durng the process that c s composed wth other components to form the whole doman models. As can be foreseen, the coarser c s granularty s, the few the tmes of composton wth other components s, and the lower CPC(c) s. Defnton (Change cost) Component c s change cost CGC(c) s defned as the cost that c adjusts ts structure or behavors to satsfy dfferent busness requrements when c s reused n multple busness envronments. In feature space, the change cost can be regarded as the cost to nstantate every feature contaned n c, or, the cost to choose a rght feature tem for every feature. Easly understood, the coarser c s granularty s, and more easly the features contaned n c changes, the hgher CGC(c) s. Change cost s closely related to the stablty property (wll be defned n secton 5.1) of features n c. In order to decrease change cost, we have to mprove the stablty of feature n c, or decrease c s granularty. And n order to adequately utlze the advantages of coarsegraned components on reuse effcency and composton cost, we may only adopt the approach of mprovng stablty of features n components wthout losng the granularty. One of the most mportant tasks durng component dentfcaton s to seek the optmzed granularty so as to decrease the change cost of coarse-graned components as much as possble under to guarantee that the reuse effcency and composton cost both keep n an optmzaton level. In next secton we wll present the concept of dynamc granularty, and by acqurng components wth dynamc granulartes, we try to obtan the ntegrated optmzaton on component s reuse performance. 4 Mappng between busness model and component model As dscussed above, the process of component reuse s the process to construct doman feature space usng components. Whle a doman feature space s obtaned from a set of busness models. Here we frstly gve some basc descrptons about busness models, then dscuss the mappngs between doman feature space and component space. 4.1 Busness model and ts feature space In the doman of enterprse nformaton system, busness models reflect the basc management patterns of the enterprse, and are the foundatons for the desgn of enterprse nformaton systems. Busness models n the same doman may have some smlartes between dfferent ndustres and companes accordng to dfferent manufacturng types, such as MTO, ETO, MTS, etc. These smlar busness models consttute a doman busness model together. There exsts multple vews n busness models, and every vew contans several types of busness elements. These elements can be classfed nto statc elements and dynamc elements, the former of whch ncludes doman, organzaton, poston, busness object, etc, and the latter of whch ncludes doman flow, busness process, busness actvtes, etc, whch depct the busness flows between dfferent departments or postons n enterprses. When we use these models to construct software components, the two types of busness elements could be mapped to entty components and process components, respectvely. Mechansms for the two knds of mappngs are qute dfferent. Mappng from statc elements to entty components s relatvely easer, and n lteratures there exsts many methods whch have mplemented ths mappng, the man approach of whch s by aggregaton,.e., groupng multple closely related fne-graned elements nto a coarse-graned entty component. Mappng from dynamc elements to process components seems a lttle more complex, the man dea of whch s by decomposton,.e., clusterng a complex busness process nto some loosely-coupled sub-processes and map every one of them to a coarse-graned process component. In ths paper we prmarly dscuss the dentfcaton and desgn of process components. Accordng to ther semantcs types, elements whch consttute a busness process model can be dvded nto several layers,.e., process, sub-process, actvty, operatons, etc, and every layer s called a basc granularty layer. There exsts composton relatonshps between elements n two neghborng layers,.e., an element n an upper layer are composed of a set of elements n lower layer. Therefore, a sngle busness process forms a feature space, and can be regarded as a feature tree, n whch the busness process element s the root feature, and s decomposed gradually downwards nto sub-process features, actvty features and operaton features. Busness rules between actvtes or operatons (such as ECA rules) can be expressed by a set of FSD. In the followng, a busness

7 ACM SIGSOFT Software Engneerng Notes Page 7 May 2006 Volume 31 Number 3 element e s accordngly denoted by a feature f n doman feature space. Suppose a busness process bp forms a feature space Ω(bp), there exsts n layers L 1, L 2,, L n, and denote F as the busness elements (or features) n st layer. For the element f +1, j n layer F +1, there exsts a set of elements chld(f +1, j )={f 1, f 2,, f k n layer F that makes f +1, j = R{f 1, f 2,, f k, where R s a composton operator. 4.2 Basc desgn process of reusable components A component provdes specfc servces that support the mplementaton of one or several busness models, therefore a component can be regarded as a part of busness models. There exsts a mappng relatonshp between busness models and components, and can be expressed by the followng formula: ( Abstract( Decompose( BM ))) ( Adapt( C Aggregate (1) BM Composte Confg/ Instantate/ Select C (2) ( ))) Expresson (1) s the process of component desgn. By applyng the operatons of decomposton, abstracton and aggregaton on a set of related busness models BM, we get a set of components C. Three phases s ncluded n ths process, just as follows: Decomposton. Accordng to specfc rules, decompose busness models nto a set of sub-models, and map every sub-model nto a component. Some of decomposton rules nclude couplngcoheson rule [6], data dependency rule and functon dependency rule [2], etc. Abstracton. Abstract those smlar servces wth same busness goals and dfferent mplementaton mechansms so that a compo- nent can be reused n multple busness stuatons to mprove reusablty. Practcal abstracton technques nclude dmensonalty reducton, groupng, splttng, and ntensonalzaton, etc [2]. An abstract component can mplement multple somers of one specfc busness,.e., dfferent mplementaton styles of one bus- of ness, whch s called vertcal abstracton. Ths knd of component usually provdes dfferent busness logc or busness rules to deal wth the same busness objects. For example, n a Purchasng product arrval management component, t can realze dfferent busness logc that s conduced by the dfferent temporal order the arrval of products and nvoce. An abstract component can also mplement the common functons used n multple busnesses, whch s called horzontal abstract. Ths knd of components usually has a majorty of commonaltes and a mnorty of specfc busness logc/rules, and can deal wth dfferent busness objects. For example, n a Sales order management component, t can deal wth common orders, retal orders, long-term sales agreements, etc. Aggregaton. For those components whch are often reused together, aggregate them to get a coarse-graned component so as to mprove the reuse effcency and reuse cost. Related aggregaton technques nclude Common Reuse Prncple (CRP), nhertance/composton-based aggregaton, etc [2][23]. In component desgn process, decomposton phase specfes the basc granularty of every component; abstracton phase has no effect on component granularty, and aggregaton phase ncreases component granularty. Expresson (2) descrbes the process of component reuse, durng whch, approprate components are frstly quered and dstlled from lbrary, and by confguraton, nstantaton and modfcaton on these components, we can compose them nto a system to satsfy the requrements descrbed by busness models. (1) and (2) are two reverse processes. 4.3 Mappng strateges between busness models and component models How to create the mappng between busness models and component feature space, and decompose busness models nto a set of reusable components, s a key problem n component desgn process. Ths mappng should satsfy the propertes of completeness and non-ntersecton. Completeness means that every busness element n busness models can be mplemented by one or several components, denoted as ( ) f Ω bp, c, c,..., c C, n 1, 1 2 { k n1 n2 nk makes f =R f, f,..., f,..., f, f,..., f, ( ) n whch f PS c, 1 n, 1 j k n j Sngle granularty layer mappng (SG-mappng) In SG-Mappng, one granularty layer L s frstly specfed, then every element f belongng to layer L n doman feature space (rep- n c {f,.e., F(c {f )={f descendant(f); For resentng a specfc doman busness model) s drectly mapped to a component,.e., for f F(L ), f s mapped to a component c {f, denoted as f c {f. In addton, f s all descendant elements descendant(f) are all contaned all the elements whose layer are hgher than f, they can be mple- mented by the composton of components obtaned by SGmappng. The mappng s shown n Fg.2. n=1 means that f s a feature that s provded by a sngle component; n>1 means that f must be mplemented by the composton of features provded by multple components {c 1, c 2,, c n. Non-ntersecton means that for arbtrary two components c, c j, ther feature spaces should not ntersect wth each other, Ω(c ) Ω( c j )=,.e., there do not exst any features that appear n two or more components at the same tme. Based on the concluson above and comparsons between several exsted component desgn methods n lteratures, we classfy the mappng nto four types: SG-mappng, MG-mappng, IG-mappng and DG-mappng. L Fg.2 Sngle granularty level mappng: SG-mappng SG-mappng s smple, but ts shortcomngs are also obvous,.e., obtaned components are all n the same granularty layer, whch wll lead to poor reusablty. For example, f L s lower n doman feature space, these components are fne-graned wth lower reuse effcency and hgher composton cost, but f L s hgher, then those components are coarse-graned, but are possble to lead to poor change cost. 1 n (3)

8 ACM SIGSOFT Software Engneerng Notes Page 8 May 2006 Volume 31 Number Multple granularty layer mappng (MG-mappng) MG-mappng s based on SG-mappng. It maps elements n multple granularty layers synchronously and separately to realze mult-granularty component co-exstences. For example, for two layers L, L j (L <L j ), we apply SG-mappng separately on each element n L and L j. For each element g n the layer below L, f f s a feature n layer L and f ancestor(g), then g s contaned n the component c {f obtaned from f by SG-mappng. For each element between layer L and L j, t can be mplemented by composton of the components obtaned from SG-mappng on L. For each element n layer hgher than L j, t can be mplemented by the composton of components obtaned from SG-mappng on L j. MGmappng s shown n Fg.3. The prmary beneft of MG-mappng s that granulartes of components are more flexble, and components wth dfferent granulartes can co-exst n the result component set. But ths knd of flexblty s only lmted nto those basc granularty layers n busness models. L j L Fg.3 Multple granularty level mappng: MG-mappng Mddle granularty layer mappng (IG-mappng) A smlar problem n SG-mappng and MG-mappng s that the fnal components granularty s bascally specfed, that s to say, when a granularty layer s specfed, components to be obtaned are all n the same granularty layer of the correspondng busness elements and cannot be changed. The man dea of IG-mappng s, frst specfy one layer L, and accordng to the couplng degree between elements n ths layer, cluster these elements to get multple components. These components has granularty that s between L and L j. For example, f we specfy the busness actvty layer for IG-Mappng, every obtaned component contans one or several busness actvtes, as shown n Fg.4. L Fg.4 Mddle granularty layer mappng: IG-mappng For IG-mappng, t can also realze multple mddle granularty layer mappng (MIG-mappng).e., specfy multple layers, and cluster elements n every layers separately to form components. IG-mappng realzes some flexblty on component granularty to some extent, but the flexblty s only lmted between the specfed granularty layer and ts neghborng low layer Dynamc granularty layer mappng (DG-mappng) SG-mappng, MG-mappng and IG-mappng have the same characterstcs that the components granularty s bascally specfed when one or several granularty layers are chosen before the mappng, and the flexblty on granularty s very small. In addton, all the three mappngs do not consder the semantcs propertes of every busness element, so the granularty of components ndeed has no essental relatons wth the characterstcs of these elements, whch wll lead to poor reusablty. It s obvous that the mappng should produce components wth dynamc granularty, whch s called DG-mappng. If we combne the above three mappng strateges together, for arbtrary feature f n arbtrary layer L, when f s mapped to component space, there are three possble strateges: Drectly mapped to a component; Mapped as a part of a component Composed by a set of components Fg.5 Dynamc granularty layer mappng The key of DG-mappng s the mappng prncple,.e., accordng to whch specfed prncples to map one element to component space, and how to ensure hgh reusablty, as shown n Fg.5. We wll present a component desgn method based on busness model stablty, n whch busness elements stablty s evaluated to determne whch of the three mappng strateges s used n DG- each mappng for element. In ths secton we frstly dscuss the relatonshps between model stablty and component granularty, and then we present a metrcs for model stablty from two aspects: stablty of busness elements, stablty of relatonshps between elements. 5 Stablty and stablty dependency The nstablty of busness models s expressed n two aspects: nstablty on tme axs, nstablty between dfferent ndustres or enterprses. The former behaves that, a busness s runnng wth current pattern, but after a perod of tme, t changes to another patterns. The latter behaves that, dfferent ndustres/enterprses have dfferent patterns for a same busness. The stablty of a busness model s defned as the degree to keep stable for the busness model both n tme and n space, and can be evaluated by the de- gree of frequently changng of ths model. 5.1 Relatonshps between busness model stablty and component granularty As dscussed above, components are obtaned from busness mod- system has been els by a specfc mappng strategy. An nstable busness model needs frequent changes, but the correspondng software system that supports the runnng of ths model should not be changed so frequently. If changes happen after the software normally runnng, modfcatons on t wll have to lead to fre- quently nterruptons and restarts, whch wll also cause huge cost. Therefore, the components whch are acqured from busness models by mappng and mplement software systems by composton should have to accommodate to these changes,.e., components should be consdered as an solated area, or a cache, between nstable busness models and stable software systems n order to avod the frequent changes on software systems. Ths requres that a component tself should have the ablty of adapt to changes,.e., t can satsfy dfferent busness requrements by confguraton, and wll not or rarely nfluence the correspondng software system. If

9 ACM SIGSOFT Software Engneerng Notes Page 9 May 2006 Volume 31 Number 3 changes on busness models cannot be abstracted and encapsulated nto component n a confgurable form (.e., busness rules, parameters, etc), the granularty of components has to be decreased, and these changes wll have to be realzed n reuse process by developers for concrete busness crcumstances, otherwse, components have to face to the huge change cost of frequent changes when reused. For example, for a busness actvty element Producton plannng, ts nput s sales order nformaton, and ts output s producton plans, but accordng to dfferent manufacturng types, such as ATO, ETO, MTO, etc, the nner plannng algorthms are qute dfferent, so t has a set of somers. In Fg.6 we present an example of somers, n whch, (a), (b) and (c) descrbe three somers τ 1, τ 2, τ 3 of f, the correspondng chld feature set are {u, v, w,{u, v,{u, v, and nstantated to {u 1, v 1, w 1,{u 1, v 2,{u 2, v 3, respec- tvely. (d) shows the whole feature tree of f. From the analyss above we can know, component granularty s closely related to the stablty of busness models. The more stable the busness model s, the few changes t wll have, then, the few changes the correspondng components need to adapt to, and the coarser granularty the component wll have. So does t on the contrary,.e., the less stable the busness model s, the more changes t wll have, then the more dffcult the correspondng components adapt to these changes, so we have to decompose the busness model nto some fner-graned components. Therefore, for stable busness models, they can be drectly mapped to coarse-graned components. For nstable busness models, they should be decomposed nto a set of sub-models, map those stable sub-models nto components, and sequentally decompose those nstable sub-models, untl every obtaned sub-model are stable enough and all the busness elements n ths model have been mapped to components. 5.2 Stablty and stablty metrcs A busness element s stablty s closely related to the number of ts somers. Dfferent somers realze the same busness goals, but wth dfferent mplementaton strateges. If we regard a busness element as a feature n feature space, then the set of all ts somers s the nstance set of the correspondng feature, and every somer can be denoted as a feature nstance. For ther nner structures, dfferent somers nclude dfferent chld feature set, or same chld feature set but nstantated to dfferent feature tems. Defnton 5.1 (Isomer) A busness element f s somers are defned as the nstances of f. All of them have the same nput/output, but wth dfferent mplementaton, denoted as R(f)={τ 1, τ 2,.., τ n. τ 1 u 1 v 1 w 1 τ 2 u 1 v 2 τ 3 u 2 v 3 (a) (b) (c) Fg.6 Examples of somers u 1 f τ 1 τ 2 τ 3 v 1 u u v v 2 w w 1 2 v 3 The stablty S(f) of busness element f, s related wth the followng three factors: number of somers N(f), dstrbuton on reuse frequency of somers P(f), and smlarty between somers L(f) Number of somers The larger the number of somers N(f) s, the more frequently f (d) wll change, and the less stablty t has. The range of N(f) s [1, + ) and N(f)= nstance(f) Dstrbuton on reuse frequency of somers: stablty entropy N( f) measures the stablty of busness elements from one aspect, but the relatonshp between N(f) and S(f) s not absolutely lnear. S(f) s also related to the dstrbuton on reuse frequency of somers. We use stablty entropy to measure t. Defnton 5.2 (reuse frequency) A busness element f s one so- as ρ(τ ), s defned as the propor- mer τ s reuse frequency, denoted ton of τ s reuse tmes n f s total reuse tmes n a perod of tme, and satsfes 0 ρ(τ ) 1. ρ( τ ) could be ganed from the reuse statstcal data n a perod of tme, whch s usually 6 to 24 months, and could comparatvely reflect the degree of frequent reuse for every somers factually. Defnton 5.3 (stablty entropy) Suppose an element f s somer set s R(f)={τ 1, τ 2,..., τ n, for τ R(f), ts reuse frequency s ρ(τ ), then f s stablty entropy s denoted as P n ( f ) = ρ ( τ ) log ρ( τ ) (4) = 1 A larger P(f) ndcates that the dstrbuton on reuse frequency of dfferent somers s more balanced, and f s less stable because f has more chances to be frequently swtched between ts somers. A smaller P(f) ndcates that the dstrbuton s more lopsded, and f s more stable. The range of P(f) s [0, + ). If N(f)=1, then there must exst P(f)=0, whch ndcates that f has only one mplementaton and t s absolute stable Smlarty between somers Even f N(f) and P( f) are both very large, f s stll possble to be stable. Ths s because that stablty s also related to the smlarty between somers. Defnton 5.4 (smlarty between somers) The smlarty L(f) s defned as the degree of smlarty between the nner mplementaton of all ts somers. The more smlar between somers, the less dfference between them, and the more stablty f has. The range of L(f) s [0,1]. The metrcs to evaluate L(f) can be separately dscussed accordng to f s type: atomc element and complex element. Atomc elements refer to those elements that are n the bottom of doman feature space and cannot be decomposed any more (leaf features), e.g., busness operatons. Complex elements refer to those elements that can be decomposed nto a set of chld elements. For an atomc element, we can smply get the smlarty of ts somers by estmaton. For a complex element, ts smlarty between somers can be measured by the synthess of smlartes between every chld elements and relatonshps between these chld elements. For the latter, because relatonshps can be descrbed by busness rules to realze confgurablty, when evaluatng smlarty of complex elements, we gnore t and only consder the smlartes between chld elements. We use algorthm 1 to calculate the smlarty between somers of complex elements. In addton, although some elements are not atomc, f t s unnecessary to be decomposed or s dffcult to be decomposed, they can