26 June 2019 in IEEE/ACM Transactions on Networking
Traffic engineering (TE) is a fundamental task in networking. Conventionally, traffic can take any path connecting the source and destination. Emerging technologies such as segment routing, however, use logical paths that are composed of shortest paths going through a predetermined set of middlepoints in order to reduce the flow table overhead of TE implementation. Inspired by this, in this paper, we introduce the problem of node-constrained TE, where the traffic must go through a set of middlepoints, and study its theoretical fundamentals. We show that the general node-constrained TE that allows the traffic to take any path going through one or more middlepoints is NP-hard for directed graphs but strongly polynomial for undirected graphs, unveiling a profound dichotomy between the two cases. We also investigate a variant of node-constrained TE that uses only shortest paths between middlepoints, and prove that the problem can now be solved in weakly polynomial time for a fixed number of middlepoints, which explains why existing work focuses on this variant. Yet, if we constrain the end-to-end paths to be acyclic, the problem can become NP-hard. An important application of our work concerns flow centrality, for which we are able to derive complexity results. Furthermore, we investigate the middlepoint selection problem in general node-constrained TE. We introduce and study group flow centrality as a solution concept, and show that it is monotone but not submodular. Our work provides a thorough theoretical treatment of node-constrained TE and sheds light on the development of the emerging node-constrained TE in practice.
6 April 2020 in IEEE Transactions on Services Computing
The parameter server architecture is prevalently used for distributed deep learning. Each worker machine in a such system trains the complete model, which leads to a large amount of network data transfer between workers and servers. We empirically observe that the data transfer has a major impact on training time. We present a new distributed training system called Stanza to tackle this problem. Stanza exploits the fact that in many models such as convolution neural networks, most data exchange is attributed to the fully connected layers, while most computation is carried out in convolutional layers. Thus, we propose layer separation in distributed training: most nodes of the cluster train only the convolutional layers, while the rest train the fully connected layers. Gradients and parameters of the fully connected layers no longer need to be exchanged across the entire cluster, thereby substantially reducing the data transfer volume. We implement Stanza on PyTorch and evaluate its performance on Azure and EC2. Results show that Stanza accelerates training significantly over current parameter server systems: on EC2 instances with Tesla V100 GPU and 10Gb bandwidth for example, Stanza is 1.34x–13.9x faster for common deep learning models.