Towards integrating learning algorithms into computer system design

Machine learning brings opportunities for designing efficient computer systems by potentially identifying a variety of system patterns. The evolving learning methods provide promising approaches to improve both software and hardware systems via smart job scheduling, resource allocation, and optimized data structures. The learning methods, particularly deep learning algorithms, in turn, introduce new uncertainties while being integrated into computer systems due to the lack of interpretability and robustness. In this thesis, we look into specific problems of the two aspects. First, how to use learning theories to improve the efficiency of computer systems. We particularly look into the data deduplication systems for both primary and backup storage as cases; Second, while using deep learning algorithms, how to improve the interpretability and the robustness of the models so that we are more confident to integrate them into systems. Correspondingly, this thesis consists of the following two parts. In Part I, we aim to use the statistical learning methods to solve two fundamental problems in deduplication systems, namely the cache efficiency problem in primary deduplication systems and the sub-optimal configuration problem in backup deduplication systems. We show that using a unified unseen estimation method is able to solve both of the problems elegantly. In primary storage deduplication, using the estimation approach enables us to infer the locality of workloads so that the fingerprint cache can achieve a much higher hit rate. For backup storage, the accurate estimations of chunking benefits are useful to guide the selection of better chunking configurations. In Part II, we investigate the critical issues preventing the adoption of deep learning algorithms in critical systems. Particularly, we study the convolutional neural networks and the graph convolutional networks. We explore two research problems: First, how can we provide the users with an intuitive way to understand the decision boundaries of a convolutional neural network? We propose methods to construct data structures to assist the interpretations; Second, how will the adversarial users attack the graph convolutional networks and how can we defend? We present an integrated gradients guided method for stronger attacks and a pre-processing based method as a defense