ALIC: A Low Overhead Compiler Optimization Prediction Model

Iterative compilation based on machine learning can automatically predict the best optimization for the new programs. However, the efficient prediction models often require repetitive training, which leads to a higher training time overheads, and greatly affects the widespread utilization of the technology. The existing approaches in the prediction model construction often use random sample search strategy, which easily lead to data redundancy. In addition, due to the effect of run-time noises, the sample program is subjected to a fixed number of repetitive observations. However, in the case there is very little noises, the repetitive observations will result in a serious waste of iterative compilation time overheads. Therefore, how to effectively collect the optimal prediction model samples and choose the appropriate sample observations number are the key problem of reducing the iterative compilation overheads. We propose a low overheads iterative compilation optimization parameters prediction model ALIC. First, we describe the target programs by static-dynamic features representation based on feature-class relevance, and construct an initial optimization prediction model by the classifier. Then we use a dynamic number of sample observations strategy for each sample. The most profitable sample from the candidate samples set is selected and marked, each mark is equivalent to increase the number of sample observations. Finally, the optimization prediction model is constructed based on the intermediate prediction model that learns candidate samples actively. The experimental results show that when predicting optimization parameters for the new programs on Intel Xeon E5520 and Chinese Shenwei 26010 platforms, the ALIC model generates 1.38× (by ICC14.0 compiler), 1.35× (by GCC5.4 compiler) average performance improvement on the Xeon platform, and 1.42× (by SW compiler) on the Shenwei Platform. In addition, the ALIC model can significantly reduce the iterative compilation training time overheads than the existing approaches.