Optimizing Machine Learning on Apache Spark in HPC Environments

Machine learning has established itself as a powerful tool for the construction of decision making models and algorithms through the use of statistical techniques on training data. However, a significant impediment to its progress is the time spent training and improving the accuracy of these models - this is a data and compute intensive process, which can often take days, weeks or even months to complete. A common approach to accelerate this process is to employ the use of multiple machines simultaneously, a trait shared with the field of High Performance Computing (HPC) and its clusters. However, existing distributed frameworks for data analytics and machine learning are designed for commodity servers, which do not realize the full potential of a HPC cluster, and thus denies the effective use of a readily available and potentially useful resource. In this work we adapt the application of Apache Spark, a distributed data-flow framework, to support the use of machine learning in HPC environments for the purposes of machine learning. There are inherent challenges to using Spark in this context; memory management, communication costs and synchronization overheads all pose challenges to its efficiency. To this end we introduce: (i) the application of MapRDD, a fine grained distributed data representation; (ii) a task-based all-reduce implementation; and (iii) a new asynchronous Stochastic Gradient Descent (SGD) algorithm using non-blocking all-reduce. We demonstrate up to a 2.6x overall speedup (or a 11.2x theoretical speedup with a Nvidia K80 graphics card), a 82-91 % compute ratio, and a 80% reduction in the memory usage, when training the GoogLeNet model to classify 10% of the ImageNet dataset on a 32-node cluster. We also demonstrate a comparable convergence rate using the new asynchronous SGD with respect to the synchronous method. With increasing use of accelerator cards, larger cluster computers and deeper neural network models, we predict a 2x further speedup (i.e. 22.4x accumulated speedup) is obtainable with the new asynchronous SGD algorithm on heterogeneous clusters.