Training jobs of deep neural networks (DNNs) can be accelerated through distributed variants of stochastic gradient descent (SGD), where multiple nodes process training examples and exchange updates. The total throughput of the nodes depends not only on their computing power, but also on their networking speeds and coordination mechanism (synchronous or asynchronous, centralized or decentralized), since communication bottlenecks and stragglers can result in sublinear scaling when additional nodes are provisioned. In this paper, we propose two classes of performance models to predict throughput of distributed SGD:fine-grained models, representing many elementary computation/communication operations and their dependencies; andcoarse-grained models, where SGD steps at each node are represented as a sequence of high-level phases without parallelism between computation and communication. Using a PyTorch implementation, real-world DNN models and different cloud environments, our experimental evaluation illustrates that, while fine-grained models are more accurate and can be easily adapted to new variants of distributed SGD, coarse-grained models can provide similarly accurate predictions when augmented with ad hoc heuristics, and their parameters can be estimated with profiling information that is easier to collect.

Modern machine learning frameworks can train neural networks using multiple nodes in parallel, each computing parameter updates with stochastic gradient descent (SGD) and sharing them asynchronously through a central parameter server. Due to communication overhead and bottlenecks, the total throughput of SGD updates in a cluster scales sublinearly, saturating as the number of nodes increases. In this paper, we present a solution to predicting training throughput from profiling traces collected from a single-node configuration. Our approach is able to model the interaction of multiple nodes and the scheduling of concurrent transmissions between the parameter server and each node. By accounting for the dependencies between received parts and pending computations, we predict overlaps between computation and communication and generate synthetic execution traces for configurations with multiple nodes. We validate our approach on TensorFlow training jobs for popular image classification neural networks, on AWS and on our in-house cluster, using nodes equipped with GPUs or only with CPUs. We also investigate the effects of data transmission policies used in TensorFlow and the accuracy of our approach when combined with optimizations of the transmission schedule.