Software defect prediction plays a vital role to identify the most defect prone modules or components of software. Its aim is to determine software reliability via learning from historical defect data. Feature selection is used to train the prediction models. It helps to enhance the performance of prediction and reduce the computation time of models. Various studies have been carried out on feature selection methods within the project or cross projects.  The purpose of this work is to synthesize the literature of previous studies on  different feature selection techniques  with respect to small size of metrics data in different context ,  to find out  the choice of training data,  and the modelling techniques applied that have great  impact on the performance of  prediction models. We have conducted the literature review to evaluate the feature selection methods proposed from 2009 to 2019.The results are analysed qualitatively and quantitatively of  24 studies. It reveals approaches and adequate context-specific used information based on the criteria we construct and apply. The discussions are identified from  24 studies  and analysed by considering the assessment points. As per the literature review feature selection techniques performed well , contributed to allocate minimal and relevant metrics data and reduced the computation time. Thus the feature selection is core step for any classifier to improve the overall prediction output.