Accelerating Iterative Computations for Large-Scale Data Processing

Recent advances in sensing, storage, and networking technologies are creating massive amounts of data at an unprecedented scale and pace. Large-scale data processing is commonly leveraged to make sense of these data, which will enable companies, governments, and organizations, to make better decisions and bring convenience to our daily life. However, the massive amount of data involved makes it challenging to perform data processing in a timely manner. On the one hand, huge volumes of data might not even fit into the disk of a single machine. On the other hand, data mining and machine learning algorithms, which are usually involved in large-scale data processing, typically require time-consuming iterative computations. Therefore, it is imperative to efficiently perform iterative computations on large computer clusters or cloud using highly-parallel and shared-nothing distributed systems. This research aims to explore new forms of iterative computations that reduce unnecessary computations so as to accelerate large-scale data processing in a distributed environment. We propose the iterative computation transformation for well-known data mining and machine learning algorithms, such as expectation-maximization, nonnegative matrix factorization, belief propagation, and graph algorithms (e.g., PageRank). These algorithms have been used in a wide range of application domains. First, we show how to accelerate expectation-maximization algorithms with frequent updates in a distributed environment. Then, we illustrate the way of efficiently scaling distributed nonnegative matrix factorization with block-wise updates. Next, our approach of scaling distributed belief propagation with prioritized block updates is presented. Last, we illustrate how to efficiently perform distributed incremental computation on evolving graphs. We will elaborate how to implement these transformed iterative computations on existing distributed programming models such as the MapReduce-based model, as well as develop new scalable and efficient distributed programming models and frameworks when necessary. The goal of these supporting distributed frameworks is to lift the burden of the programmers in specifying transformation of iterative computations and communication mechanisms, and automatically optimize the execution of the computation. Our techniques are evaluated extensively to demonstrate their efficiency. While the techniques we propose are in the context of specific algorithms, they address the challenges commonly faced in many other algorithms