Pam McDonald Dan Strickland Charles Wildman THAAD Project Office Dynetics, INC. THAAD Project Office

Transcription

1 Estimating the Effective Size of Auto-Generated Code in a Large Software Project Pam McDonald Dan Strickland Charles Wildman THAAD Project Office Dynetics, INC. THAAD Project Office DISTRIBUTION A: Approved for public release; distribution unlimited.

2 Overview Background Terminology Nature of of Testing Inputs Testing vs. vs. Non-Testing Effort Synthetic Estimate Cases Additional Areas of of Interest

3 Background Massive amounts of of software can can be be generated by by code generation languages or or tools Graphical User Interfaces (GUIs) or or high-order languages can can generate large amounts of of code saving time and and effort Auto-generated code counts can can be be over-inflated making inaccurate counts for for size, effort, productivity, and and defect density Parametric models like like COCOMO II II and and SEER-SEM don t have the the ability to to account for for auto-generated code sizes GOAL: Find a solution that that correctly estimates size/effort of of Auto-Generated Code

4 Terminology generating related to to the the development, input effort used to to autogenerate code resultant related to to the the code that that is is output by by the the auto-generation synthetic the the normalized estimate of of effort that that takes into account the the effort in in non-testing activities (generating language) and and testing activities (resultant language)

5 Nature of Testing Determine the the type of of testing to to be be performed on on the the code No No testing on on resultant code Product constrained by by time-to-market issues Testing to to be be performed on on generating language only only Uses Uses generating estimate as as estimate for for size size in in cost cost model No No calculation of of synthetic effort Testing on on resultant code required Product is is mission critical with with potential loss loss of of life life risks risks Non-testing activities are are performed in in the the generating language Testing activities are are performed in in the the resultant language Synthetic effort must must be be calculated for for size size input input to to cost cost models

6 Inputs Generating Language (e.g. Visual Basic, IDL, PowerBuilder) Resultant Language (e.g. Ada95, C++, FORTRAN) Generating Estimate (in (in SLOC or or a measure convertible to to SLOC) Resultant Estimate (in (in SLOC or or a measure convertible to to SLOC) Function Point Estimate (one estimate works for for generating and and resultant languages)

7 Testing vs. Non-Testing Efforts COCOMO II Phases Waterfall Model Requirements COCOMO II Effort Distribution Large Project Prelim. Design Detailed Design IT 26.2% RQ 6.5% PD 15.9% Non-Testing Testing Code and Unit Testing Integration Testing CUT 29.0% DD 22.5% Requirements, Preliminary Design and and Detailed Design are are nontest related; Integration Testing is is test test related Code and and Unit Testing contains activities in in both areas test

8 Testing vs. Non-Testing Efforts - Code and Unit Testing Phase Code and Unit Testing Requirements 4% Design 8% Programming 56.5% Test Plans 5.5% V & V 8.5% Project Office 6% CM/QA 6.5% Manuals 5% 77.25% Non- Test 22.75% Test Non-Testing Testing Code and and Unit Unit Test Test broken up up into into default activities Requirements, Design, and and Programming are are non-test activities Test Test Plans and and V & V are are test test activities Project Office, CM/QA, and and Manuals are are non-specified --split split evenly between both both efforts

9 Testing vs. Non-Testing Effort - Synthetic SLOC COCOMO II Effort Distribution Non-Testing Testing IT 26.2% RQ 6.5% PD 15.9% 67.2 % Non-Test CUT (test) 6.6% CUT (non) 22.4% DD 22.4% 32.8% Test Normalized effort effort values values for for non-test non-test and and test test efforts efforts 67.2% 67.2% of of development size size counted counted in in synthetic size size 32.8% 32.8% of of translated size size counted counted in in synthetic size size Example: 10KSLOC translated to to 50KSLOC 10,000 10,000 (0.672) (0.672) + 50,000 50,000 (0.328) (0.328) = 23,120 23,120 SLOC SLOC Synthetic SLOC SLOC Rule Rule of of Thumb: 66.7% 66.7% non-test, 33.3% 33.3% test test

10 Using the Programming Languages Table SLOC Estimates Average Language Statements per FP Ada95 49 C 128 C++ 53 ZLISP 64 Function Point Estimates Programming Languages Table (Capers Jones) allows for for a rough order of of magnitude conversion between SLOC and and Function Points The table allows for for taxation of of language based on on a common denominator (Function Points)

11 Synthetic Estimate Case 1 Inputs: Generating Language Known, Resultant Language Known, Generating Estimate in in SLOC Formula: Synthetic SLOC = (66.7%) * (Generating SLOC) + (33.3%) * (Resultant Language PLT PLT Value // Generating Language PLT PLT Value) * (Generating SLOC) Example: SLOC of of IDL IDL will will be be used used to to auto-generate a software item item in in Ada95. What size size input input should be be used? Synthetic SLOC = (66.7%) * (5,000) + (33.3%) * (49/20) * (5000) Synthetic SLOC = Synthetic SLOC =

12 Synthetic Estimate Case 2 Inputs: Generating Language Known, Resultant Language Known, Resultant Estimate in in SLOC Formula: Synthetic SLOC = (66.7%) * (Generating Language PLT PLT // Resultant Language PLT) * (Resultant SLOC) + (33.3%) * (Resultant SLOC) Example: SLOC of of C will will be be auto-generated using PowerBuilder. What size size input input should be be used? Synthetic SLOC = (66.7%) * (16/128) * (50000) + (33.3%) * (50000) Synthetic SLOC = Synthetic SLOC = 20819

13 Synthetic Estimate Case 3 Inputs: Generating Language Known, Resultant Language Known, Estimate in in Function Points Formula: Synthetic SLOC = (66.7%) * (Generating Language PLT PLT * Function Point Estimate) + (33.3%) * (Resultant Language PLT PLT * Function Point Estimate) Example: A 4GL 4GL auto-generation language will will be be used used to to produce a software item item in in C++. C++. There are are function points estimated. What size size input input should be be used? Synthetic SLOC = (66.7%) * (20) (20)* (450) + (33.3%) * (53) (53)* (450) Synthetic SLOC = Synthetic SLOC = 13946

14 Synthetic Estimate Case 4 Inputs: Generating Estimate in in SLOC, Resultant Estimate in in SLOC Formula: Synthetic SLOC = (66.7%) * (Generating SLOC) + (33.3%) * (Resultant SLOC) Example: An An estimated statements in in IDL IDL auto-generate Ada95 SLOC. What size size input input should be be used? Synthetic SLOC = (66.7%) * (3500) + (33.3%) * (50000) Synthetic SLOC = Synthetic SLOC = 18985

15 Adapted Code Adapted code is is pre-existing code modified to to be be reused with new development Modification percentage is is made up up of of percent re-design, percent recoding, and and percent re-testing/integration factors If If the the adapted code is is in in the the same language as as the the generating language, more attention must be be paid to to the the percent re-design and and re- percent re-coding factors If If the the adapted code is is in in the the same language as as the the resultant language, more attention must be be paid to to the the percent re-test and and integration factor Adapted SLOC is is added to to Synthetic SLOC and and any any new, non-auto generated SLOC for for use use as as the the size size input when reuse exists Size = Adapted SLOC + Synthetic SLOC + New [Non-AGC]

16 Maintenance COCOMO II II Maintenance Model: Major size size input input is is Total Delivered SLOC (TSLOC) Major modifier is is Annual Change Traffic Percentage When there is is auto-generated code, the the TSLOC is is at at least as as large as as the the resultant SLOC, potentially inflating the the estimate Annual Change Traffic is is driven by by changes to to code in in both generating and and resultant languages Changes to to generating language mean complete regeneration and and retesting of of the the code, inflating the the change traffic Changes to to resultant language could be be lengthy and and time-consuming given the the size-to-effort ratio and and potential unfamiliarity with the the code Continued research is is needed in in this topic

17 Conclusion The use of code generators is becoming more prevalent in large software projects, but parametric models don t have the ability to properly model the effort of auto-generation Using generating and resultant languages, specific inputs, and testing vs. non-testing effort percentages allows for the calculation of a Synthetic size to be used for effort calculations Further research is needed into how auto-generated code affects maintenance Using the methods described here, auto-generated code effort can be accounted for in parametric estimates