Medicamentsen-ligne vous propose les traitements dont vous avez besoin afin de prendre soin de votre santé sexuelle. Avec plus de 6 ans d'expérience et plus de 90.000 clients francophones, nous étions la première clinique fournissant du acheter viagra original en France à vente en ligne et le premier vendeur en ligne de Viagra dans le monde. Pourquoi prendre des risques si vous pouvez être sûr avec Medicamentsen-ligne - Le service auquel vous pouvez faire confiance.

Cist.buct.edu.cn

Evaluating C++ Design Pattern Miner Tools Lajos F¨ul¨op, Tam´as Gyovai and Rudolf Ferenc University of Szeged, Department of Software Engineering {flajos|ferenc}@inf.u-szeged.hu very important that these definitions were informal, becausethis way a pattern can be used in a wider context. Due to Many articles and tools have been proposed over the the imprecise pattern definitions the implemented structures years for mining design patterns from source code. These based on them could vary in different contexts. Design pat- tools differ in several aspects, thus their fair comparison is terns also differ in the aspect that they are used intentionally hard. Besides the basic methodology, the main differences or only in a casual way. A programmer can apply a design are that the tools operate on different representations of the pattern, without actually knowing about it.
subject system and that the pattern definitions differ as well. Design patterns are employed in many areas of software In this paper we first provide a common measurement development. Originally, their main aim was to develop bet- platform for three well-known pattern mining systems, ter software systems by using good solutions. Another good Columbus, Maisa and CrocoPat. Then we compare these property of design patterns is, that their documentation in a tools on four C++ open-source systems: DC++, WinMerge, software system can simplify maintenance and program un- Jikes and Mozilla. Columbus can discover patterns from the derstanding. This is especially true for large software sys- C++ source code itself, while Maisa and CrocoPat require tems. Unfortunately, the developers usually do not provide the representation of a software system in a special textual this documentation, so there is a big need to discover design format, so we extended Columbus to provide the common patterns from source code. Therefore, in the past years var- ious design pattern miner tools have been developed. These We compared these tools in terms of speed, memory con- sumption and the differences between the hits. While the One of the important aspects is the programming lan- first two aspects showed comparable results, the recogni- guage. There are tools for discovering patterns from Java tion capabilities were quite diverse. This is probably due to source code like Ptidej [13, 21] and Fujaba [11], and tools the significant difference in how the patterns to be recog- also exist for mining patterns from C++ code like Colum- nized and formalized by the tools. Therefore we conclude bus [2, 9]. Certain tools like CrocoPat [3] and Maisa [20] that a more precise and formal description of design pat- work based on special own textual format which describe facts about the source code needed to find design patterns.
These tools are general, because they can work on any pro-gramming language, only the appropriate textual input has to be prepared from the source code.
Another aspect is the method used to discover design pat- Design pattern mining, Tool evaluation, Columbus, terns, which can be quite diverse. Columbus uses graph matching, while Maisa solves constraint satisfaction prob-lems (CSP). CrocoPat has a new method to find structures in large graphs, it makes an effective representation of rela-tions in graphs. Other special and interesting methods also Design patterns are well-known structures in the soft- exist, like pattern inference, which was presented by Tonella and well-tried solutions for common recurring problems.
Our aim in this article was to compare three design pat- Gamma et. al. [12] collected several object oriented design tern miner tools: Columbus, Maisa and CrocoPat. We chose patterns, and they gave an informal definition for them. It is these tools, because it was possible to prepare a common input for them with our front end, Columbus. Our previous pressive thus it is able to describe design patterns, design work enabled us to provide the input for Maisa [10], while metrics or other structures. In section 3.2 we will present in case of CrocoPat we created a new plug-in for Columbus which is able to prepare the appropriate input. Finally, the Arcelli et. al. [1] proposed three categories for design tools have been compared in three aspects: differences be- pattern mining tools, considering the information which tween the hits, speed and memory requirements. We think was used during the detection process. These categories are that these are the most important aspects in a design pattern the “entire” representation of design patterns, the minimal miner tool. We did not analyze if a found design pattern hit set of key structures that a design pattern consists of, and the is true or false, we examined these tools only with the con- sub-components of design patterns. They have dealt with sideration of structural hits and differences in this aspect.
the last category, and two tools concerning this, FUJABA We will proceed as follows. In Section 2, we will discuss and SPQR were introduced and compared. The base of the some works similar to ours. In Section 3 we will introduce comparison was how a tool decomposes a design pattern the design pattern miner tools, which were compared. Sec- into smaller pieces. The conclusion was, that the decompo- tion 4 describes our comparison approach, and our results sition methods of the two examined systems are very simi- are presented in Section 5. Finally, in Section 6 we will lar, and finally they argued the benefits of sub-patterns.
present some conclusions and outline directions for future Gu´eh´eneuc et. al. [14] introduced a comparative frame- work for design recovery tools. The purpose of the authors’framework was not to rank the tools but to compare them with qualitative aspects. This framework contained eightaspects: Context, Intent, Users, Input, Technique, Output, In this section we show some similar works to ours, and Implementation and Tool. These aspects were sorted into 53 we also present new and interesting methods in the area of criteria which were demonstrated on two systems, Ptidej and LiCoR. The major need for this framework is that, al- In our previous work we have presented a method to dif- though there are a lot of design recovery tools, the compari- ferentiate true and false hits [8]. We employed machine son between them is very hard due to the fact that they have learning methods to filter out false design pattern hits. First, very different characteristics in terms of representation, out- we ran our design pattern miner tool that discovers patterns put format and implementation techniques. This framework based on structural descriptions. Afterwards, we classified provides an opportunity for comparing not only similar sys- these hits as being true or false, and finally we calculated tems, but also systems, which are different. We note that predictive information for the hits. We trained a decision this comparison differs from ours because we compare sys- tree based on classified values and on the predictive infor- tems in a practical way. Namely, we want to show and com- mation, from which we were able to mine true design pat- pare how many patterns a tool can find, and how much time and memory it needs for searching, while their comparison Design pattern detection was also accomplished by the is rather theoretical, it does not compare the discovering ef- integration of two existing tools – Columbus [9] and Maisa [20] – in our previous work [10]. This method com- Kaczor et. al. [16] proposed a bit-vector algorithm for bined the extraction capabilities of the Columbus reverse design pattern identification. The algorithm initialization engineering system with the pattern mining ability of Maisa.
step converts the design pattern motif and the analyzed pro- First, the C++ code was analyzed by Columbus. Then the gram model into strings. To model the design patterns facts collected were exported to a clause-based design nota- and the analyzed program, six possible relations can be tion understandable for Maisa. Afterwards, this file was an- used between elements: association, aggregation, compo- alyzed by Maisa, and instances were searched that matched sition, instantiation, inheritance and dummy. The authors the previously given design pattern descriptions. Maisa ap- gave an efficient Iterative Bit-vector Algorithm to match proached the recognition problem as a constraint satisfac- the string representation of the design patterns and the tion problem. We will get back to this method in Sec- analyzed program. They compared their implementation with explanation-based constraint programming and metric- Beyer et. al. [3, 4] have developed a system that is able to enhanced constraint programming approaches.
work with large graphs effectively. The effectiveness of the Costagliola et. al. [6] based their approach on a visual system is based on binary decision diagrams which repre- language parsing technique. The design pattern recogni- sent the relations compactly. They have developed the rela- tion was reduced to recognizing sub-sentences in a class tion manipulation language (RML) for manipulating n-ary diagram, where each sub-sentence corresponds to a design relations and a tool implementation (CrocoPat) that is an pattern specified by an XPG grammar. Their process con- interpreter for the language. The RML language is very ex- sist of two phases: the input source code is translated into a class diagram represented in SVG format; then DPRE (De- to Prolog. Figure 2 shows the description of the Factory sign Pattern Recovery Environment) recovers design pat- terns using an efficient LR-based parsing approach.
Tonella and Antoniol [22] presented an interesting ap- proach to recognize design patterns. They did not use a class(”Product”).
class(”ConcreteProduct”).
library of design patterns as others did but, instead, discov- extends(”ConcreteProduct”,”Product”).
ered recurrent patterns directly from the source code. They !same(Product,ConcreteProduct).
class(Creator).
employed concept analysis to recognize groups of classes method(”Creator.FactoryMethod()”).
sharing common relations. The reason for adapting this ap- has(”Creator”,”Creator.FactoryMethod()”).
returns(”Creator.FactoryMethod()”,”Product”).
proach was that a design pattern could be considered as a formal concept. They used inductive context construction extends(”ConcreteCreator”,”Creator”).
method(”ConcreteCreator.FactoryMethod()”).
which then helped them to find the best concept.
has(”ConcreteCreator”,”ConcreteCreator.FactoryMethod()”).
creates(”ConcreteCreator.FactoryMethod()”,”ConcreteProduct”).
implements(”ConcreteCreator.FactoryMethod()”,”Creator.FactoryMethod()”).
returns(”ConcreteCreator.FactoryMethod()”,”Product”).
!same(Creator,ConcreteCreator).
!same(Product,Creator).
In this study we compared the design pattern mining capabilities of three tools, namely Maisa, CrocoPat and Columbus. Maisa and CrocoPat cannot analyze sourcecode, so we extended our Columbus framework (whoseoriginal task was to analyze C++ source code and build Figure 2. Factory Method pattern in Maisa
an ASG – Abstract Semantic Graph representation fromit) to produce input files for the two tools. So, our studywas based on the same input facts, this way ensuring afair-minded comparison, because eventual parsing errors af- fected all tools in the same way. We illustrate this processin Figure 1.
In the next sections we will introduce the compared In Section 2 we already introduced the CrocoPat tool [3] design pattern miner tools. We will show every tools’ briefly, and now we will describe it in detail. First we will design pattern description language on the well-known show how the CrocoPat interpreter works, and then we will introduce the relational manipulation language. Finally wewill also mention the binary decision diagrams (BDD).
Previously, we mentioned that CrocoPat is an interpreter, and it executes RML programs. First, CrocoPat reads the Maisa is a software tool [20] for the analysis of software graph representation in rigi standard file format (RSF) [19] architectures developed in a research project at the Univer- from the standard input. Afterwards, the RML description sity of Helsinki. The key idea in Maisa is to analyze design is processed and a BDD representation is created from it.
level UML diagrams and compute architectural metrics for Finally, the RML program is executed and an RSF output is early quality prediction of a software system.
In addition to calculating traditional (object-oriented) software metrics such as the Number of Public Methods, The RML (Relational Manipulation Language) is very Maisa looks for instances of design patterns (either generic similar to logic programming languages like Prolog, but it ones such as the well-known GoF patterns or user-defined contains techniques of imperative programming languages special ones) from the UML diagrams representing the soft- too. Hence, it is very expressive and it can describe design ware architecture. Maisa also incorporates metrics from dif- patterns among other structures. Unfortunately, we have not ferent types of UML diagrams and execution time estima- found any design pattern library in RML, so we had to cre- tion through extended activity diagrams.
ate the descriptions of the patterns by ourselves. Figure 3 Maisa uses constraint satisfaction [17], which is a shows the Factory Method description in CrocoPat.
generic technique that can be applied to a wide variety of To sum up, the main goals of Beyer et. al. [4] was effi- tasks, in this case to mining patterns from software archi- ciency and easy integration with other tools when they de- tectures or software code. A constraint satisfaction problem veloped CrocoPat. Integration was facilitated by the im- (CSP) is given as a set of variables and a set of constraints port and export of relations in the simple Rigi Standard For- restricting the values that can be assigned to those variables.
mat (RSF), and efficiency was achieved by representing the Maisa’s design pattern description language is very similar relations as binary decision diagrams [5].
Figure 1. Common framework
AbstractClass(X) := CLASS(X) & ABSTRACT(X); (plug-ins) of the system. Some of these plug-ins are pro- vided as basic parts of Columbus, while the system can be ConcreteProduct(Cpr,Pr) := CLASS(Cpr) & Product(Pr) & extended to meet other reverse engineering requirements aswell. This way we have got a versatile and easily extendible Creator(Cr,Pr) := AbstractClass(Cr) & ASSOCIATION(Cr,Pr) & Product(Pr) & CreatMethods(Cr,Pr,M) := Creator(Cr,Pr) & HASMETHOD(Cr,M);CreatorFM(Cr,Pr,FM) := CreatMethods(Cr,Pr,FM) & VIRTUAL(FM) & One of the plug-ins is CAN2Dpm, which discovers design patterns. The design patterns were described in CreatorAM(Cr,Pr,AM,FM) := CreatMethods(Cr,Pr,AM) & DPML (Design Pattern Markup Language) files, whichstore information about the structures of the design patterns.
ConcreteCreator(Ccr,Pr,Cr,Cpr) := CLASS(Ccr) & ASSOCIATION(Ccr,Pr) & Product(Pr) & Creator(Cr,Pr) & TC(INHERITANCE(Ccr,Cr)) & CAN2Dpm recognizes design patterns in the following way.
First, Columbus analyzes the source code and builds an Ab- CCreatorFM(Ccr,Pr,Cpr,M) := ConcreteCreator(Ccr,Pr, ,Cpr) & HASMETHOD(Ccr,M) & VIRTUAL(M) & !PUREVIRTUAL(M) & stract Semantic Graph (ASG) that contains all the informa- tion about the source code. Then CAN2Dpm loads a DPML FactoryMethod(Prod,Creat,CProd,CCreat,CreatFM,CreatAM,CcreatFM) := file which also basically describes a graph. Afterwards it tries to match this graph to the ASG using our algorithm Creator(Creat,Prod) &ConcreteProduct(CProd,Prod) & described in previous work [2]. Figure 4 shows the Factory ConcreteCreator(CCreat,Prod,Creat,CProd) & CreatorFM(Creat,Prod,CreatFM) &CreatorAM(Creat,Prod,CreatAM,CreatFM) & The other two plug-ins of Columbus shown in Figure 1, CCreatorFM(CCreat,Prod,CProd,CcreatFM) & CAN2Maisa and CAN2CrocoPat, are responsible for creat- ing the input files for Maisa and CrocoPat, respectively.
Figure 3. Factory Method pattern in CrocoPat
In this section we will present the comparison approach of the investigated design pattern mining tools concerning: • The found design pattern instances. Differences are Columbus is a reverse engineering framework, which has caused by several reasons. The different tools use dif- been developed in cooperation between FrontEndART Ltd., ferent techniques to define and describe what is a de- the University of Szeged and the Software Technology Lab- sign pattern. The recognition algorithms are also dif- oratory of Nokia Research Center. Columbus is able to an- ferent. The comparison of the found design pattern alyze large C/C++ projects and to extract facts from them.
instances is just one of the several kinds of evaluation The main motivation to develop the Columbus system has that should be considered, therefore we measure other been to create a general framework to combine a number of important characteristics like speed and memory.
reverse engineering tasks and to provide a common inter-face for them. Thus, Columbus is a framework tool which • Speed. Speed is measured by the amount of the time supports project handling, data extraction, data representa- taken by the tool to perform the selected design pattern tion, data storage, filtering and visualization. All these ba- mining on the selected C++ project. The differences sic tasks of the reverse engineering process for the specific in the time of the measuring process in the examined needs are accomplished by using the appropriate modules systems are described in Section 5.2.
<?xml version=’1.0’?> cation domain. These four real-life, freely available C++ <!DOCTYPE DesignPattern SYSTEM ’dpml-1.6.dtd’> <DesignPattern name=’Factory Method> <Class id=’id10’ name=’Creator’ isAbstract=’true’> • DC++ 0.687. Open-source client for the Direct Con- <Association ref=’id30’ /><Operation id=’id11’ name=’FactoryMethod’ kind=’normal’ nect protocol that allows to share files over the internet isVirtual=’true’ isPureVirtual=’true’> <hasTypeRep ref=’id50’/> </Operation><Operation id=’id12’ name=’AnOperation’ kind=’normal’> • WinMerge 2.4.6. Open-source visual text file differen- <calls ref=’id11’/> tiating and merging tool for Win32 platforms [23].
<hasTypeRep ref=’id54’/> • Jikes 1.22-1. Compiler that translates Java source files as defined in The Java Language Specification into the <Class id=’id20’ name=’ConcreteCreator> <Base ref=’id10’ /> byte-coded instruction set and binary format defined in <Association ref=’id30’ /> The Java Virtual Machine Specification [15].
<Operation id=’id21’ name=’FactoryMethod’ kind=’normal’ isVirtual=’true’ isPureVirtual=’false’> <creates ref=’id40’ /> • Mozilla 1.7.12. All-in-one open source Internet appli- <hasTypeRep ref=’id50’/> cation suite [18]. We used a checkout dated March 12, <Class id=’id30’ name=’Product’ isAbstract=’true’></Class> Table 1 presents some information about the analyzedprojects. The first row shows how many source and header <Class id=’id40’ name=’ConcreteProduct’ isChangeable=’true’ > <Base ref=’id30’ /> files were analyzed in the evaluated software systems. The second row lists the size of these source and header files in <TypeRep id=’id1’/> The last two rows were calculated by the metric calcu- <TypeRep id=’id50’> lator plug-in of Columbus, and gives information about the total lines of code (LOC) and the number of classes. Under <hasReturnTypeRep ref=’id52’/> the term of LOC we mean every line in source code that is not empty and is not a comment line (also known as “logical <TypeRep id=’id52’> <TypeFormerType ref=’id30’/> <TypeRep id=’id54’> <hasReturnTypeRep ref=’id56’/> <TypeRep id=’id56’> Table 1. Size information of the projects
<TypeFormerType ref=’id1’/> All tests are run on the same computer so the measured values are independent from the hardware and thus the re- Figure 4. Factory Method pattern in DPML
sults are comparable. Our test computer had a 3 GHz IntelXeon processor with 3 GB memory. In the next chapter wewill describe our benchmark results and evaluate them in • Memory usage. Memory usage is another performance measure. We measured the total memory required forthe design pattern mining task. It was complicated be- cause the memory usage of CrocoPat is fixed, and onlythe Columbus source code was available to extend it toprovide us with memory usage statistics. The applied In this section we will present our results concerning memory measuring method for the examined tools are the differences between the design pattern instances found, the running-time and the memory requirements. In thenext subsection we will start with the discovered pattern in- We did the comparison on four open source small-to- stances, and then compare the time efforts of the tools. Fi- huge systems, to make the benchmark results independent nally, we will show the memory requirements of the design from system characteristics like size, complexity and appli- Table 3. WinMerge hits
Table 2. DC++ hits
Maisa did not have this precondition. In the second case thefound pattern in Maisa had a Target participant that was not In this section we will present our experiments regard- abstract, which was a requirement in Columbus. If we re- ing pattern instances found by the compared design pattern laxed the description of this pattern in Columbus, it found miner tools. Unfortunately, Maisa did not contain descrip- these two instances too. The best solution would be if an ex- tions of the patterns Bridge, Chain of Responsibility, Deco- act definition existed for this pattern in both tools. CrocoPat rator, State, Strategy and Template Method (the results for found six Adapter Object instances, while Columbus found these are marked with dashes in our tables). We have in- only three. The cause was that if Columbus found a pattern vestigated the differences between the tools manually, so instance with certain participant classes and another pattern we checked and compared the found instances and the de- instance existed with the same participant classes but par- scriptions of design patterns in all of the cases. We will not ticipating with different methods, Columbus considered it explain every difference, because there is not enough space as being the same pattern. This is a very important differ- for it, but we will present the most common causes.
ence between Columbus and CrocoPat, so we will refer to First, we summarize our results on DC++ in Table 2.
this difference several times. Maisa found a Builder in Win- This was a small software system, so it did not contain too Merge but the two other tools did not, because in Maisa the many design pattern instances. Maisa found two Adapter Builder pattern representation did not contain the Director Classes, while CrocoPat and Columbus found none. This participant while the two other tools did contain it. In the is due to the fact that the definition of the Adapter Class case of State, Strategy and Template Method the differences in Maisa differed from those in Columbus and CrocoPat.
were due to that Columbus counted pattern instances partic- In Maisa the Target participant class was not abstract and ipating with different methods only once, like in the case of the Request method of the Target class was not pure vir- tual, while in Columbus and CrocoPat these features were Next, we will describe our experiments on design pat- requested. We have examined the two Adapter Class hits tern instances found in Jikes (see Table 4). Maisa found an in Maisa, and we have found that the Targets were not ab- Adapter Class, while Columbus and CrocoPat did not. The stract in these cases and the Request operations were not reason was the same as in the case of DC++, namely that in pure virtual. Columbus and CrocoPat found 14 State and 14 Maisa the Target participant class was not abstract and the Strategy design pattern instances. The cause of the identical Request method of the Target class was not pure virtual but number of hits is that the State and Strategy patterns share in Columbus and CrocoPat these features were required. In the same static structure, so their description in the tools the case of Adapter Object Maisa missed a lot of hits, while Columbus and CrocoPat could discover a lot of design pat- Table 3 shows the results of the tools in the case of Win- tern instances. It looked like CrocoPat found more instances Merge. Maisa found two more Adapter Objects in Win- because Columbus counted repeating patterns with different Merge than Columbus. In the first case the difference was operations only once. Actually, these tools found the same caused by the fact that the Request method of a participant pattern instances. In Maisa the Builder pattern representa- Adapter Object class was defined virtual in Columbus while tion did not contain the Director participant, so Maisa found lot of design pattern instances were found, like in the case of State, where CrocoPat found 7662 and Columbus discov- ered 722 instances. This huge difference was due to the fact that the found design pattern instances were not grouped by CrocoPat, that is, if a design pattern contained a class with child classes where the child classes could be of arbitrary number, every repeated child class with the common par- ent appeared as a new hit. Columbus recognized this situa- tion and handled it correctly. In the case of Adapter Class the causes of differences were the same as in Jikes and in DC++ examined earlier. Columbus did not count repeated instances in the case of Adapter Object, so it actually found the same instances as CrocoPat, but Maisa missed some because of its different pattern description. CrocoPat andColumbus found the same instances of the Bridge patternbut Columbus counted the repeating patterns with different Table 4. Jikes hits
operations only once. Maisa found false Builder instancesagain, because the description of this pattern did not contain the Director participant class. Maisa found Factory Meth- ods instances while the two other tools did not. This is due to that the two other tools defined Factory Method with an abstract Product and an abstract Creator participant class, while Maisa did not require these participants to be abstract.
CrocoPat did not find any Mediator instance in Mozilla, while Maisa discovered two instances. This is due to that Maisa described Mediator in a very special way, so it con- tained a Mediator with two Colleagues, but Concrete Me- diators were missing. In the case of Prototype, Singleton, State, Strategy and Template Method the differences were caused again by that CrocoPat counted every repeated pat- tern instance while Columbus counted these repeated ones with different operations only once.
Because of space limitation we cannot explain every dif- ference, but we have shown the common reasons. Basically, Table 5. Mozilla hits
the found design pattern instances would be the same inmost of the cases if we could disregard the following com-mon causes of differences: an incomplete Builder instance in Jikes. CrocoPat did notfind any Mediator in Jikes, while Maisa found four. It is due • Different definitions of design patterns. We have found to that Maisa described Mediator in a very special way, so that there were some specific reasons for that the tools that it contained a Mediator with two Colleagues, but Con- discovered different pattern instances. The main rea- crete Mediators were missed. The description of Mediator son was in some cases that a design pattern description in CrocoPat required a Mediator abstract class with a child missed a participant like in the case of the Builder pat- ConcreteMediator class, too. In the case of Proxy, every tern in Maisa. In this case the pattern definition did not tool discovered 53 instances, but CrocoPat counted also re- contain the director participant, thus the instances dis- peating patterns with different methods. Maisa found 21 covered by Maisa differed from the results of the other instances more because it did not require an abstract Proxy tools. For example, the results of Maisa in WinMerge participant class in the Proxy design pattern. In the case of for the Builder pattern differed from those of CrocoPat State and Strategy it seemed that Columbus found less de- sign pattern instances but it counted every repeated patternwith different methods only once. Maisa found 23 Visitor • Precision of pattern descriptions. Another difference patterns, that the two other tools did not. This is due to the was how precise and strict the pattern descriptions loose description of this pattern in Maisa.
were. For example, in the case of Jikes the differences Table 5 shows our experiments in the case of Mozilla. A in the numbers of found Adapter Class instances were caused by the fact that CrocoPat and Columbus defined the Target as abstract while Maisa did not.
• Differences in algorithms. We have perceived dif- ferences in the design pattern miner algorithms, too.
Columbus and Maisa counted the repeated instances with different operations only once while CrocoPat In this section we will present and compare the speed performance of the three assessed design pattern miner tools. We wanted to measure only the search time for pat- terns, therefore we divided the running time into two parts,an initialization part and a pattern mining part. Tables 6, 7, 8and 9 contain the values of the pattern mining time only.
Table 6. DC++ times
Table 10 contains the initialization time of the tools (timeformat: hh:mm:ss).
The design pattern mining time was measured in the fol- • Columbus. We took into account only the graph match- ing time, so we did not consider the time while the ASG was loaded. The graph loading time is presented • CrocoPat. In the case of CrocoPat, we have prepared a small tool which executed CrocoPat and measured its running time. We measured the time needed for every pattern mining procedure for every subject soft- ware system. Next, we also measured the time for the subject systems with an empty RML program, because this way we could measure the time necessary to re-serve the memory and to prepare the BDD represen- Table 7. WinMerge times
tation (initialization time). These results are shown inTable 10. Finally, we subtracted the initialization timefrom the full running time for every result, and this The time requirements for discovering patterns in Win- way obtained the pattern matching times.
Merge (see Table 7) and Jikes (see Table 8) were differ- • Maisa. Maisa created statistics for every pattern min- ent. Columbus was very fast in the case of larger patterns, ing procedure, which contained information about the because it could filter out [2] a lot of class candidates at time necessary for pattern mining, so we used these the beginning of the discovering process. Opposite to this, generated statistics. Contrary to CrocoPat, there was Columbus was slower in the case of smaller patterns, be- no need to extract the initialization time, because time cause in these cases a lot of class candidates remained for values in the generated statistics measured only the the detailed discovering process. CrocoPat’s and Maisa’s pattern mining phase. However, we also show the ini- time requirements were very balanced.
tialization time for Maisa in Table 10.
Finally, Table 9 shows the results for Mozilla. In most cases, CrocoPat delivered the best results, but in certain First we show our results for DC++ (see Table 6). In cases Columbus and Maisa were faster. Columbus was slow this case the required time was very small for every as- when it could filter out only a small amount of class candi- sessed pattern miner tool, therefore they can be considered dates at the beginning of the discovering process. The CSP as being equal. This is due to the small size of the DC++ algorithm of Maisa was also slow in this case.
system, hence the design pattern instances were discovered Our conclusion was that the best tool regarding speed in very quickly in this system by all three tools.
general is CrocoPat, but in some cases Columbus was faster.
Table 10. Initialization times
• CrocoPat. CrocoPat’s memory usage is constant and can be set as a command line parameter. Therefore, we created a script that executed CrocoPat iteratively from 1 megabyte reserved memory up to 200 megabytes for every pattern mining process. We took the smallest possible value so that the pattern mining process stillcompleted successfully.
Table 8. Jikes times
Our experiment proved that the memory usage strongly depended on the size of the analyzed projects and it was in- dependent from the searched design patterns. This was true for every pattern miner tool as it can be seen in Table 11.
Table 11. Memory requirements in megabytes
In the case of Columbus the reserved memory was very large compared to the other tools. This is due to the fact Table 9. Mozilla times
that Columbus is a general reverse engineering frameworkand design pattern detection is only one of its many fea- tures. For this reason it uses an ASG representation, whichcontains all information about the source code (includingdetailed facts about statements and expressions not needed In this section we will introduce and compare the mem- for design pattern detection) for all kinds of tasks. Right ory usage of the three compared design pattern miner tools.
now, for technical reasons, the design pattern miner plug-in We have measured the memory requirements of every de- of Columbus does not work without the ASG (although it sign pattern mining procedure, but we show our results sum- does not have to use it), but we wish to fix this in the future.
marized in one table because we have found them very sim- Therefore, we measured the memory needed by Columbus also without the ASG and showed these numbers in paren- The memory measurement method in the examined sys- tems was accomplished in the following way: Note, in the case of CrocoPat and Maisa the reserved • Columbus. We have extended the tool, so that it reports memory was smaller because their input contained only the information about the source code necessary for pattern de-tection.
• Maisa. Maisa did not report the memory usage in its After examining Table 11 we can conclude that in the as- statistics, so we measured it by simply monitoring its pect of memory requirement Maisa’s performance was the peak memory usage on the task manager.
[6] G. Costagliola, A. D. Lucia, V. Deufemia, C. Gravino, and M. Risi. Design Pattern Recovery by Visual Language Pars- In this paper we have presented a comparison of three ing. In Proceedings of the 9th Conference on Software design pattern miner tools: Columbus, Maisa and CrocoPat.
Maintenance and Reengineering (CSMR’05), pages 102– We have compared them regarding patterns hits, speed and 111. IEEE Computer Society, Mar. 2005.
memory consumption. We have guaranteed the common in- put for the tools by analyzing the source code with the front http://sourceforge.net/projects/dcplusplus/ end of Columbus and by creating plug-ins for producing the A. Besz´edes, L. F¨ul¨op, and J. Lelle. Design Pat- required files for the tools. This way, as a “side effect” of tern Mining Enhanced by Machine Learning. In Proceed- this work, we have extended our Columbus Reverse Engi- ings of the 21th International Conference on Software Main-tenance (ICSM 2005), pages 295–304. IEEE Computer So- neering Framework with plug-ins for Maisa and CrocoPat.
We conclude that the fastest tool is CrocoPat, and Maisa re- A. Besz´edes, M. Tarkiainen, and T. Gyim´othy.
quires the least memory, while Columbus is an all-in-one Columbus – Reverse Engineering Tool and Schema for C++.
solution for design pattern detection from C++ source code In Proceedings of the 18th International Conference on with comparable performance to the other two specialized Software Maintenance (ICSM 2002), pages 172–181. IEEE Originally, Gamma et. al. [12] defined the design pat- [10] R. Ferenc, J. Gustafsson, L. M¨uller, and J. Paakki. Recog- terns to develop object-oriented applications in forward en- nizing Design Patterns in C++ programs with the integration gineering. Therefore, pattern definitions were informal to of Columbus and Maisa. Acta Cybernetica, 15:669–682,2002.
make them easier to use in different languages and con- texts. Consequently, the design pattern miner tools have http://www.cs.uni-paderborn.de/cs/fujaba/ a big problem in common, which is how to define a given [12] E. Gamma, R. Helm, R. Johnson, and J. Vlissides. Design design pattern. In this paper we have shown that the tools Patterns : Elements of Reusable Object-Oriented Software.
found different design pattern instances in common inputs mostly because of their different pattern definitions. Hence, [13] Y.-G. Gu´eh´eneuc and N. Jussien. Using explanations for de- a formal description of design patterns is very desirable.
sign patterns identification. In Proceedings of IJCAI Work- In the future we plan to create a design pattern catalog shop on Modelling and Solving Problems with Constraints,pages 57–64, Aug. 2001.
for reverse engineers, where every design pattern will have [14] Y.-G. Gu´eh´eneuc, K. Mens, and R. Wuyts. A Comparative a strict and formal description. With this new collection Framework for Design Recovery Tools. In Proceedings of of design patterns the presented drawbacks in Section 5.1 the 10th Conference on Software Maintenance and Reengi- (different definitions of design patterns, precision of pattern neering(CSMR’06), pages 123–134. IEEE Computer Soci- descriptions and differences in methods) can be avoided.
[15] IBM Jikes Project. http://jikes.sourceforge.net/ [16] O. Kaczor, Y.-G. Gu´eh´eneuc, and S. Hamel. Efficient Iden- tification of Design Patterns with Bit-vector Algorithm. InConference on Software Maintenance and Reengineering [1] F. Arcelli, S. Masiero, C. Raibulet, and F. Tisato. A Compar- (CSMR’06), pages 175–184. IEEE Computer Society, 2006.
ison of Reverse Engineering Tools based on Design Pattern [17] A. K. Mackworth. The logic of constraint satisfaction. Artif. Decomposition. In Proceedings of the 15th Australian Soft- Intell., 58(1-3):3–20, 1992.
ware Engineering Conference (ASWEC’05), pages 677–691.
[18] The Mozilla Homepage. http://www.mozilla.org/ [19] H. A. M¨uller, K. Wong, and S. R. Tilley. Understanding [2] Z. Balanyi and R. Ferenc. Mining Design Patterns from C++ Software Systems Using Reverse Engineering Technology.
Source Code. In Proceedings of the 19th International Con- In Proceedings of ACFAS, 1994.
ference on Software Maintenance (ICSM 2003), pages 305– [20] J. Paakki, A. Karhinen, J. Gustafsson, L. Nenonen, and 314. IEEE Computer Society, Sept. 2003.
A. Verkamo. Software Metrics by Architectural Pattern [3] D. Beyer and C. Lewerentz. CrocoPat: Efficient pattern Mining. In Proceedings of the International Conference on analysis in object-oriented programs. In Proceedings of the Software: Theory and Practice (16th IFIP World Computer 11th IEEE International Workshop on Program Comprehen- Congress)., pages 325–332, 2000.
sion (IWPC 2003), pages 294–295. IEEE Computer Society, [4] D. Beyer, A. Noack, and C. Lewerentz. Efficient Rela- [22] P. Tonella and G. Antoniol. Object oriented design pat- tional Calculation for Software Analysis. In Transactions tern inference. In Proceedings of the International Confer- on Software Engineering (TSE’05), pages 137–149. IEEE ence on Software Maintenance (ICSM ’99), pages 230–238, Washington, DC, USA, 1999. IEEE Computer Society.
[5] R. Bryant. Graph-Based Algorithms for Boolean Function Manipulation. In Transactions on Computers, pages 677– http://sourceforge.net/projects/winmerge/ 691. IEEE Computer Society, Feb. 1986.

Source: http://cist.buct.edu.cn/zheng/scam/2006/papers/26.pdf

2011 oct rough

1 0 2 2 3 S a w m i l l P k w y P o w e l l O H 4 3 0 6 5 We have reached October, our tenth month of our 12 month program of fitness. Our goal this month is to buy a new pair of shoes. A tennis shoe lasts 300-500 miles or at least twice a year get them re-placed. We have had many patients be more fit and reports of 18-32 lbs. lost, by participating in one of the many forms of lifestyle

Microsoft word - registration form.docx

PATIENT’S NAME_________________________ AGE_______ DATE OF BIRTH___________ EXPLAIN BRIEFLY WHAT SYMPTOMS BRING YOU TO THIS OFFICE: ARE ANY OF YOUR PRESENT PROBLEMS DUE TO INJURY? Yes____, No___ ARE YOU RIGHT-HANDED [ ] OR LEFT-HANDED [ ]? PAST MEDICAL HISTORY: 1. HAVE YOU EVER HAD: (Check the appropriate boxes and list year to the right) 2. PLEASE LIST IN CHRONOLOGICAL ORDER ALL HOSP

Copyright © 2010-2014 Pharmacy Pills Pdf