Graphlab: A new framework for parallel machine learning

Designing and implementing efficient, provably correct parallel machine learning (ML) algo-rithms is challenging. Existing high-level par-allel abstractions like MapReduce are insuf-ficiently expressive while low-level tools like MPI and Pthreads leave ML experts repeatedly solving the same design challenges. By tar-geting common patterns in ML, we developed GraphLab, which improves upon abstractions like MapReduce by compactly expressing asyn-chronous iterative algorithms with sparse com-putational dependencies while ensuring data con-sistency and achieving a high degree of parallel performance. We demonstrate the expressiveness of the GraphLab framework by designing and implementing parallel versions of belief propaga-tion, Gibbs sampling, Co-EM, Lasso and Com-pressed Sensing. We show that using GraphLab we can achieve excellent parallel performance on large scale real-world problems.