GenPerf

Introduction

GenPerf is a symbolic regression tool that generates synthetic data for evaluating transfer learning approaches. At a high level, GenPerf receives a source performance model that has been specified with the SPLConqueror format. GenPerf then generates genetic mutations of the source model to create a target performance model that has some similarities to the source model. The similarities are based on the core insights in our empirical study based on several criteria:

Similarity in terms of correlation of source and target response
Similarity in terms of number of influential options
Similarity in terms of number of option interactions
Similarity in terms of performance distribution

# goal
targetNumberOfInteractions = 10
targetNumberOfIndividualOptions = 5
numberOfNegativFeatures = 3
numberOfAbsolutCoefficientsAbove80 = 5
targetCorrelationHigh = 0.8
targetCorrelationLow = 0.2

Generating source performance model using Thor

Thor is a genetic generator for realistic attributed performance models. GenPerf uses Thor to generate a realistic source performance model using measurements of real world configurable systems. The detailed process of generating source performance model has been described here. Few examples of generated performance models for configurable systems such as LLVM and Lrzip are as follows:

LLVM: 207 * time_passes + 16 * gvn + 16 * licm + 12 * instcombine + 14 * inline + 3,5 * time_passes * Num1 * Num2 + 5,5 * gvn * licm * Num1 * Num1 + -3,7 * instcombine * inline * Num2

Lrzip: 43838 * level + 2218747 * compressionZpaq + 288311 * compressionLrzip + 191662 * compressionBzip2 + 34718 * compressionGzip + 11946 * encryption + 6676 * compression + 3433850 * compressionZpaq * level9 + 836940 * compressionLrzip * level8 + 720098 * compressionLrzip * level7 + 3415670 * compressionZpaq * level8 + 485719 * compressionLrzip * level9 + -1597534 * compressionZpaq * level1 + -1597084 * compressionZpaq * level3 + -1596575 * compressionZpaq * level2 + 111344 * compressionGzip * level9 + 102375 * compressionGzip * level8 + 59973 * compressionGzip * level7 + -129840 * compressionLrzip * level2 + -128920 * compressionLrzip * level1 + 42831 * compressionGzip * level6 + 21313 * compressionGzip * level5 + -55078 * compressionLrzip * level3 + 43656 * compressionLrzip * level6 + -37020 * compressionBzip2 * level1 + 3,5 * Num1 * Num2 + 4 * Num3 + 5 * Num4 * Num4

Setting the genetic mutation parameters

# mutation properties
probabilityOfMutatingCoefficient = 0.3
probabilityOfAddingFeature = 0.1
probabilityOfRemovingFeature = 0.1
probabilityOfAddingInteraction = 0.1
probabilityOfRemovingInteraction = 0.1
probabilityOfPairWiseInteraction = 0.8
probabilityOfThreeWiseInteraction = 0.2
probabilityOfSwitchingSign = 0.05
probabilityOfUniformCrossover = 0.1
probabilityOfBreeding = 0.5
standardDeviationForMutation = 10

Mutation operators

GenPerf searches the space of performance influence models to find the model that best fits the specified hypotheses in terms of similarities of the source and the target. Mutation is a genetic operator used to alter the source model from one generation of a population to the next. GenPerf implements the following operators, however, it can be easily extended by changing mutate method:

Add/remove interaction
Add/remove option
Change the coefficient for options and interactions
Sign switch

Usage

1- Specify the source model

perf_model_txt = "2*o1 + 3*o1*o2 + 4*o2"
perf_model = genModelfromString(perf_model_txt)
startingModel = Model(perf_model)
startingModel.name = "test"

2- Mutate

allModels = [copy.deepcopy(startingModel) for i in range(popsize)]
best, bestFitness, bestHistory, allModels = genetic_algorithm(allModels, startingModel, numberIterations)
print(best)
print(bestFitness)

3- Generate data

n = 1000
xTest = np.random.randint(2, size=(n, ndim))
evaluate2csv(best, xTest)

4- Run

python genmodel.py

Contact

If you notice a bug, want to request a feature, or have a question or feedback, please send an email to the tool maintainers:

pooyanjamshidi / GenPerf