There is an increased interest in building data analytics frameworks with advanced algebraic capabilities both in industry and academia. Many of these frameworks, e.g., TensorFlow, implement their compute-intensive primitives in two flavors-as multi-thread routines for multi-core CPUs and as highly-parallel kernels executed on GPU. Stochastic gradient descent (SGD) is the most popular optimization method for model training implemented extensively on modern data analytics platforms. While the data-intensive properties of SGD are well-known, there is an intense debate on which of the many SGD variants is better in practice. In this paper, we perform a comprehensive experimental study of parallel SGD for training machine learning models. We consider the impact of three factors - computing architecture (multi-core CPU or GPU), synchronous or asynchronous model updates, and data sparsity - on three measures-hardware efficiency, statistical efficiency, and time to convergence. We draw several interesting findings from our experiments with logistic regression (LR), support vector machines (SVM), and deep neural nets (MLP) on five real datasets. As expected, GPU always outperforms parallel CPU for synchronous SGD. The gap is, however, only 2-5X for simple models, and below 7X even for fully-connected deep nets. For asynchronous SGD, CPU is undoubtedly the optimal solution, outperforming GPU in time to convergence even when the GPU has a speedup of 10X or more. The choice between synchronous GPU and asynchronous CPU is not straightforward and depends on the task and the characteristics of the data. Thus, CPU should not be easily discarded for machine learning workloads. We hope that our insights provide a useful guide for applying parallel SGD in practice and - more importantly - choosing the appropriate computing architecture

Existing data analytics systems have approached predictive model training exclusively from a data-parallel perspective. Data examples are partitioned to multiple workers and training is executed concurrently over different partitions, under various synchronization policies that emphasize speedup or convergence. Since models with millions and even billions of features become increasingly common nowadays, model management becomes an equally important task for effective training. In this paper, we present a general framework for parallelizing stochastic optimization algorithms over massive models that cannot fit in memory. We extend the lock-free HOGWILD!-family of algorithms to disk-resident models by vertically partitioning the model offline and asynchronously updating the resulting partitions online. Unlike HOGWILD!, concurrent requests to the common model are minimized by a preemptive push-based sharing mechanism that reduces the number of disk accesses. Experimental results on real and synthetic datasets show that the proposed framework achieves improved convergence over HOGWILD! and is the only solution scalable to massive models.