Microsoft word - cbr97a.doc

Application of Case Based Reasoning to Legacy System Migration
J. Grimson, B. Wu, J. Bisbal, and D. Lawless 1. Introduction
Systems migration is now a major issue in established IT departments, [Bene95, Wins94].
Typically the information systems of these organisations are large, old, mission critical, and have
minimal, if any, documentation. These information systems define what we today call legacy
information systems
. The problems these systems pose to their host organisations are numerous and
important :
• the systems cannot evolve to provide new functionalities required by the organisation• they run on obsolete hardware which is expensive to maintain and reduces productivity • maintenance is expensive, tracing failures is costly and time consuming due to the lack of documentation and a general lack of understanding of the internal workings • integration efforts are greatly hampered by the absence of clean interfaces There is thus an urgent need to provide tools, methodologies and techniques not only for accessingthe data which is locked in these closed systems, but also for providing a strategy which will allowthe migration of the systems to new platforms and architectures. The exigency of this requirement isall the more highlighted by the pending "Year 2000 problem" which will render almost all legacysystems unusable.
Following currently accepted methods for legacy system migration, few efforts have been successful, [Fing96]. Current methodologies, [Brod95, Rena96, Gant95, Till95], are generallycharacterised by a series of complicated steps. The complexity of individual steps derives from theirgenerality as opposed to any specific tasks1. This leaves migration engineers constantly asking thequestion, “But how do I do that for my situation ?”. There is no facility to learn from previouslysuccessful solutions.
In this paper we propose that case based reasoning, CBR, can play a very constructive role in legacy system migration. In the following section we briefly describe CBR and it’s proposed role inlegacy system migration. Any migration project involves both the migration of the legacyapplications and the legacy data. In section 3 we present a more detailed discussion of the applicationof CBR in the migration of legacy applications. In section 4 we investigate the possible applicationsof CBR in the legacy data migration process. In the final section we present our conclusions andbriefly describe our intentions for future work in our MILESTONE2 project.
† To whom correspondence should be addressed1 One step from Brodies Chicken Little strategy is “Incrementally migrate Legacy Applications”, [Brod95].
2 MILESTONE is a collaborative project involving Trinity College Dublin, Broadcom Éireann Research,Telecom Éireann, and Ericssons which aims to provide a migration methodology supported by a generic tool-kitto aid migration engineers in the process of migrating legacy information systems to open (and/or distributed)systems.
2. Case Based Reasoning for Legacy System Migration
Case based reasoning, [Wats94, Kolo93, and Aamo96], has enjoyed widespread popularity as an alternative to expert systems for solution of problems of a certain type. CBR is useful in situationswhere a well understood model for the solution of a particular problem is unavailable. In thesesituations experts may be able to suggest solutions based on previous experience but would find itdifficult to explain why such solutions are suitable or to formalise a useful set of rules for the solutionof these problems in general. In these situations it is often more useful to compare a new problem topreviously encountered problems, select a previously successful solution to a similar problem and tomodify it to fit the new problem. This new solution can then be added to the set of previouslysuccessful solutions for future use.
We assert that the problem of legacy system migration is one which would benefit from the application of a CBR approach. In particular we are focussing on those issues that are central to allmigration projects; understanding the source code of legacy applications and understanding thestructure of legacy data. As can be appreciated it is very difficult to arrive at a simple genericmethodology or set of rules that should be followed for all cases. In practice migration engineerslearn over time to recognise patterns or features within legacy components. These features have inthe past led them to adopt, (perhaps following trial and error), a particular course of action whichresults in a satisfactory solution. A CBR system could provide a framework to support this learningby experience process. In the following two sections we investigate in more detail the application ofCBR to program understanding and data structure understanding.
3. CBR and Program Understanding
Program understanding is still a major stumbling block in the migration process. In our work we are investigating the potential of employing an Ontology in the creation of useful caserepresentations of legacy code fragments. Such an Ontology would be used to model domain specificknowledge about standard programming routines. For example a bubble sort might be modelledwithin the Ontology as a relationship between the concept of pointers, the concept of array traversaland the concept of pointer comparison, a Quicksort might be modelled as a relationship between theconcept of recursion, the concept of array partitioning and the concept of comparisons. A standard taxcalculation could be modelled as a relationship between the concept of a selection (of both anallowance and a tax rate), the concept of a subtraction (of the selected allowance), and the concept ofa multiplication (based upon a selected rate) and the concepts of allowances and tax rates. A simpleontology which could be used to model this calculation is shown in Figure 1. In this diagram conceptsare represented by boxes and relationships such as is selected by, is multiplied by, is a result of etc.
are represented as arrows. Each of these concepts would have links to triggers which would appear asvarious coding constructions in various programming languages, for example in C the concept of a pointer comparison could be triggered by: *p < *q or A[p] < A[q] an array partitioning could be triggered by: an allowance could be triggered by the literal a tax rate could be triggered by the literal We propose to represent each case in the case base as a collection of features, each feature being a concept from the ontology. When classifying a segment of code, the code will be parsed andits constructions mapped to concepts within the ontology, this code segment can then be classified asbeing of the type of programming routine modelled in the ontology with which it shares mostfeatures. Depending on the situation a standard template for such a routine may be retrieved or themost similar previously classified code segment may be retrieved. As the set of features used torepresent each case are concepts used to characterise each of the standard programming routinesmodelled in the ontology we predict that this representation will afford the classification algorithm apowerful discrimination ability, placing new cases strongly into one of the categories of the standardprogramming routines. The likely drawback to this approach is that cases which do not conform to theprogramming routines modelled in the ontology will be classified incorrectly or not at all. Onepossible solution would be to periodically revise the ontology to include models for new programmingroutines. Such a revision would be prompted by the appearance of poorly classified cases whichclearly do not conform to any of the models in the ontology.
These factors are directly related to the completeness of the ontology and the level of detail modelled within it. In related work on the use of ontologies for concept based information retrieval ithas been suggested that the trade-off between modelling effort and classification ability for ontologiesshould be decided relative to the particular problem at hand. A useful indicator of the correctgranularity of an ontology may be an empirical measure of its ability to solve the task for which it wascreated.
“Firstly, what level of granularity should an ontology be at, in other words how detailed shouldit be for a given task application? Clearly no ontology can make all possible distinctionsbetween a pair of concepts. The answer lies in the precise characteristics of the languageengineering task which is to be solved and in the relationship of the taxonomy to thattask.The effect of this on retrieval performance can be determined empirically although thisdoes not indicate the correct level of granularity in an absolute fashion”, [OSul95].
Just as no ontology can make all possible distinctions between a pair of concepts, neither can any ontology hope to capture the universe of all possible concepts. The completeness of an ontologyshould also be determined relative to its ability to perform the task for which it was created. The taskfor which our ontology of programming routines is intended may vary over time (as a result of newmigration efforts on unforseen data), therefore the only feasible method for ensuring the completenessand correct granularity of this ontology is to revise it whenever the performance of the system fallsbelow a certain threshold.
In effect, our approach attempts to embody some program understanding capabilities within a case representation. There have been a number of other approaches to program understanding such asconstraint satisfaction [Wood96], however we are unaware of any attempts to combine Case BasedReasoning and program understanding in order to facilitate the sophisticated classification of programsegments while building on previously classified examples.
3 In Ireland 48.1% and 28.1% are standard tax rates and IRL3660 and IRL7563 are standard tax free allowances Fig 1 : A simple Ontology representing a Tax Calculation Routine 4. CBR in Database Reverse Engineering
The process of reverse engineering legacy databases involves designing a target schema which is conceptually equivalent to the legacy schema. This has proven to be a very difficult processto model. For the purposes of this discussion we will divide the process into the followingcomponents : - Extracting the data structure from the legacy system- Conceptualising this physical/logical schema to an equivalent conceptual schema- Forward engineering from the conceptual schema to a new physical schema Each of these steps can benefit from the use of case based reasoning.
The degree of difficulty involved in the data structure extraction process depends greatly on the data model of the legacy system. If a database management system is used, be it relational,network, or hierarchical, the process can be as simple as analysing the Data Definition Language,DDL, and examining the data dictionary. Unfortunately the situation is much more complex for themore prevalent standard file format, for which no computerised descriptions of structure exists.
Individual source programs must be examined in order to detect partial structures of the files. Veryoften data structures are hidden (refer to Fig 2), optimisation constructs are introduced (e.g. paddingfor address alignment or record splitting when the size is greater than the page size), andspecifications are left to be procedurally encoded.
It can, as a result, be very difficult to recover the original data structure. However, by examiningsource code statements, (e.g. the variables used to fill the “filler” variable in Fig 2), the structure canbe inferred. Any padding can be discovered by knowing how large a page is and examing what isplaced in the “filler” variables.
Besides structure hiding the code can also give clues to constraints in the legacy data structure which should be extracted and enforced in the target schema. Referential integrity,cardinality of relationships, uniqueness constraints and secondary identifiers are just some suchconstraints that are left to be enforced in a procedural manner. The structure of these procedures aregenerally both simple and standard, and can thus be easily recognised, refer to fig 3.
Over time patterns of legacy code and system specific variables can be identified as relating to the underlying data structure. By recording these features, possibly in an Ontology as described insection 3, or simply as they are, individual cases can be built up over time. These cases can then beused by the migration engineer in the future to alleviate the complexity of the extraction process.
read Supplier(Name = X) into Supplierif found(S) then Part.Part_No := .
Part.Color := .
Part.Sup := Supplier.Name
Fig 3 : Example of Referential Integrity in Source Code 4.2 Conceptualisation of the extracted data structure The extracted data structure is typically still in a form that is a poor semantic reflection of its original conceptual schema. In this phase a number of transformations are performed to rediscoverthe full semantics of the legacy database. The resulting conceptual schema is typically expressed inan Entity Relationship model. A transformation is grossly speaking the process of replacing a datastructure with another one which is semantically equivalent to the former. Examples oftransformations include the replacement of one-to-many relationship types by reference attributes, theintroduction/replacement of multivalue or compound attributes, the transformation of entities intoattributes, etc. Transformations can be carried out for one of two reasons. Firstly to removeconstructs that were designed for optimisation purposes, e.g. merge/split records, denormalisation,derived variables, etc., refer to fig 4, and secondly, to arrive at the most expressive, simple, andreadable conceptual schema, refer to fig 5.
Fig 5 - Simplify by introducing of multivalued attributes The biggest problem in the conceptualisation process is the fact there is no 1-1 mapping betweenconceptual structures and their physical/logical compliant translation. A conceptual construct can betranslated into several physical/logical structures, while several different conceptual constructs can betranslated into the same physical/logical structure. We believe a CBR system could be employed todecide which transformations to apply and in what order to apply them. Features of a schema aresimple to illicit (eg a transitive functional dependency), and can thus be readily compared to pre-recorded features of previous cases. The entire process can be made considerably easier by CASEtools such as DB-MAIN, [Hain96], which graphically illustrate schemas and automatically apply anynumber of transformations to a schema.
4.3 Forward engineering from the conceptual schema to a new physical schema This step is the reverse of the conceptualisation process, i.e. employing reversible transformations to derive a physical schema, (generally in the Object Oriented data model butpossibly the Relational model), which is semantically equivalent to the conceptual schema. There isonce again an excellent opportunity to use a CBR system to aid the migration engineer in choosingwhich transformations to apply and in what order to apply them.
5. Conclusions
In this paper we have highlighted the crucial role CBR can play in the process of legacy system migration. The nature of the problem, i.e. the absence of a usable model, makes it particularlyamenable to the solution by experience approach proposed by CBR. We have concentrated on theparticularly difficult issues of legacy application understanding and the legacy data understanding.
Other approaches to this problem to date have been so complex as to preclude their application in apractical generic solution.
MILESTONE is an ongoing project and the proposals introduced in this paper are the current focus of attention. The project is working with real life legacy systems in Telecom Éireann, the Irishnational telecommunications service provider. Immediate future work includes the construction ofdomain dependent Ontologies and further investigations into possible case representations for legacydata structures. There are a number of ontologies available at present which may serve as startingpoints in the construction of domain specific ontologies. The Princeton Wordnet, [Mill90a, Mill90b],is an excellent general purpose concept ontology which may be used to relate programming specificconcepts to real-world concepts at a high level. The Enterprise Ontology, [Usch95], is a collection ofterms and definitions relevant to businesses which may provide a useful high-level framework for theorganisation of programming routines specific to certain business domains. The Knowledge SystemsLaboratory, [KSL94], within the Department of Computer Science at Stanford University provide alibrary of shareable ontologies available over the internet. Some of these ontologies such as‘Abstract-Algebra’, ‘Basic-Matrix-Algebra’ and ‘Component-Assemblies’ could be exploited within adomain specific ontology of programming routines.
[Aamo96] : Aamodt A. and Plaza E., “Case-Based Reasoning: foundational Issues, MethodologicalVariations, and System Approaches”, Artifical Intelligence Communications, Vol. 7, No.1, 1996.
[Bene95] : Bennett K., “Legacy Systems: Coping with Success”, IEEE Software, January 1995, pp.
19 - 22.
[Brod95] : Brodie M. and Stonebraker M., “Migrating Legacy Systems Gateways, Interfaces and theIncremental Approach”, Morgan Kaufmann, 1995.
[Fing96] : Fingar P., and Stikleather J., “Distributed Objects For Business”, SunWorld Online, April1996.
[Gant95] : Ganti N., Brayman W., “Transition of Legacy Systems to a Distributed Architecture”, John Wiley & Sons Inc. 1995.
[Hain96] : Hainaut J-L., Englebert V., Henrad J., Hick J-M., and Roland D., “Database DesignRecovery”, Proc. of the 8th Conference on Advance Information Systems Engineering, CAiSE’96,Springer-Verlag, 463-480.
[Kolo93] : Kolodner J. L., “Case Based Reasoning”, Morgan Kaufmann.
[KSL94] : “Shareable Ontologies Library” at [Mill90a] : George A. Miller, Richard Beckwith, Christiane Felbaum, Derek Gross, and KatherineMiller, "Introduction to WordNet : An On-line Lexical Database", International Journal ofLexicography, Vol. 3, No. 4, 1990, 235 - 244.
[Mill90b] : George A. Miller, "Nouns in WordNet : A Lexical Inheritance System", InternationalJournal of Lexicography, Vol. 3, No. 4, 1990, 245 - 264.
[Osul95] : O’Sullivan, D., McElligott, A., Sutcliffe, R.F.E, ‘Augmenting the Princeton WordNetwith a Domain Specific Ontology’ in Proceedings of the IJCAI ‘95 Workshop on Basic OntologicalIssues in Knowledge Sharing 19-21 August, 1995, Montreal, Canada.
[Rena96] : “The RENAISSANCE Project”, [Till95] : Tilley S., “Perspectives on Legacy System Reengineering”, [Usch95] : M. Uschold, M. King, S. Moralee and Y. Zorgios, “The Enterprise Ontology”, at, 1995.
[Wats94] : Watson I. and Marir, F., “Case-Based Reasoning: A Review”, The KnowledgeEngineering Review, Vol. 9, No. 4, 1994.
[Wins94] : Winsberg P., “What about Legacy Systems”, Database Programming and Design, Vol. 7,No. 3, 1994.
[Wood96] : Wood S. and Yang Q., “The Program Understanding Problem: Analysis and a HeuristicApproach”, Second Working Conference on Software Engineering, March, 1996, 78-84.



Indian Res. J. Genet. & Biotech. 5(3) : 160-164 (2013) Antibiotic sensitivity profile of the test isolates of E. coli isolated from clinical cases of diarrhea in the kids of goat breeds S. K. Rajput1, K. Gururaj2, D.V. Singh3 and Gitam Singh4 1, 2, 4 C. I. R. G, Makhdoom, Farah, Mathura, U.P., 3K.V.K., Nabarangpur, Odisha. (Received : May, 2013 : Revised : June, 2013; Accepted : J

Microsoft word - 2003-approaches.doc

Research Approaches towards a Cure for Duchenne Muscular Dystrophy A report on the state of international research for the development of a causal therapy of Duchenne muscular dystrophy Updated in August 2003 Written in cooperation with medical specialists and scientists for the German Duchenne Parent Project Aktion Benni & Co and dedicated to all Duchenne boys

© 2010-2018 Modern Medicine