This talk does not have an abstract.

Load balancing has become an important yet challenging performance optimization technology for datacenters with the immense proliferation of cloud based applications. The two dominant technical challenges when designing a datacenter load balancing mechanism are: determining the granularity of a data unit for scheduling, and determining the range of the load information to be collected. On one hand, the per-packet load balancing policies can occupy more paths for transmitting a flow, however they are also prone to cause packet reordering and degrade the performance. On the other hand, the global congestion-aware scheduling mechanisms outperform some local congestion-aware or some congestion-oblivious mechanisms by employing holistic views of the global load information; however the long duration control loops also limit the scalability of a global congestion-aware mechanism. Addressing the two issues, this paper presents a novel global congestion-aware per-packet load balancing scheduling mechanism, named SPLB (Stand-in Per-packet Load Balancing). Based on a dual channel architecture and the preview detection results, SPLB successfully maintains the arrival sequence of the data packets even under asymmetric datacenter topologies. Moreover, SPLB immediately determines the transmission route of each data packet based on the backhaul stand-in packets, without introducing any control loop durations. Extensive experiments validate the effectiveness of SPLB with selected datacenter topologies.

Accurate traffic demand matrix (TM) prediction is essential for effective traffic engineering and network management. Based on our analysis of real traffic traces from Wide Area Network (WAN), the traffic flows in TM are both time-varying (i.e. with intra-flow dependencies) and correlated with each other (i.e. with inter-flow correlations). However, existing works in TM prediction ignore inter-flow correlations. In this work, we propose a new model for TM prediction by considering both inter-flow correlation and intra-flow dependencies, namely CRNN model. By observing that most strongly-correlated traffic flows have the same source or destination, we employ convolutional structures to capture the neighboring correlation in TMs and extract even higher-order correlations. To model the trends in the time dimension, like previous works we use recurrent structures to capture the temporal variations of traffic flows. Our evaluation on two WAN datasets shows that, when predicting the next TM, CRNN model reduces the Mean Squared Error (MSE) by up to 44.8% compared to state-of-the-art method; and the gap is even larger when predicting the next multiple TMs.

There are enormous types of traffic flows in the IoT era. The traffics could be control by SDN or other transmission protocols in the core networks or mobile edge networks. TCP-BBR algorithm has a better performance on long-fat pipes than traditional TCP protocols in terms of link perception and response time. However, when TCP-BBR and other congestion control algorithms which are based on time delay like TCP-Westwood applies to the same network links, TCP-BBR will perform unfriendly to the TCP-Westwood. Aiming to improve the fairness between TCP-BBR congestion control algorithm and delay-based congestion control algorithms, this paper proposes a congestion control algorithm based on TCP-BBR which has moderate fairness named Modest Fairness BBR(MFBBR). The simulation results on the Mininet show that our algorithm can improve the fairness of BBR when it coexists with Westwood, and it also has better fairness than delay-based congestion control algorithms

With the exponential growth in mobile traffic data, enabling accurate mobile app identification is in great demand as it is an essential step to improving a multitude of network services such as QoS and security monitoring. However, the widespread use of encrypted protocols, especially the TLS protocol, has posed a challenge to the traditional traffic classification techniques for mobile app identification. With the effectiveness of the deep packet inspection approach for encrypted traffic classification being questioned, machine learning methods with statistical features have been widely studied and used. More recently, deep learning solutions are also proposed for automatic feature extraction to classify encrypted traffic. In this paper we propose App-Net, an end-to-end hybrid neural network to better learn effective features from raw TLS flows for mobile app identification. Combining RNN and CNN in a parallel way, App-Net learns a joint flow-app embedding to characterize the unique app signatures as well as flow sequence patterns, and it can be trained end-to-end from scratch to integrate both information in a unified framework. Our comprehensive experiments and comparisons with several recently proposed solutions developed based on the same real-world dataset covering 80 apps show that App-Net achieves better performance than the state-of-the-art methods.

Traffic engineering (TE) is a critical and difficult problem that involves assigning traffic with various requirements to paths with different constraints. Recently, machine learning algorithms, especially deep neural networks (DNN), are applied to TE, yet they all assume that the network is a black box, limiting them to only model-free reinforcement learning (RL) algorithms. In this paper, we introduce differentiable programming to TE, and show that the network environment can be sufficiently modeled for TE optimization. Specifically, we design a fully-differentiable network environment, dNE, that can be directly integrated into any DNN models. With dNE, we can differentiate with respect to control parameters, and directly evaluate gradients between actions and states to facilitate gradient descent based training of DNN models. We show with a proof-of-concept prototype that dNE accelerates DNN training for TE by 228x and achieves higher scalability compared to existing network simulators. Most importantly, dNE opens up the possibility to apply arbitrary deep learning models to TE beyond RL.

Network function (NF) scale out promises great flexibility and elasticity to the management of service function chains (SFCs). However, existing solutions overlook the fact that scaling out an NF may influence downstream NFs. Their scale-out plans need several steps to resolve bottleneck situations, resulting in unpredictable SFC performance and slow convergence. In this paper, we propose FastScale, a novel system that makes chain-wide scale-out plan within a single step. FastScale continuously inspects the load of each NF, i.e., realtime throughput, to accurately detect NF scale-out situations. Once a situation is detected, it models the problem of NF scale out in consideration of both target NF and downstream NFs. It determines which NFs to scale and how many NF instances should be used within a single step, which avoids unnecessary reconfiguration and reduces convergence time. After making scale-out plan, FastScale automatically updates the thresholds used to detect NF scale-out situations so as to reduce user burdens. We have built a prototype of FastScale. Our evaluation results demonstrate that FastScale reduces the convergence time of SFCs by up to 68% in comparison with existing solutions.

With the rapidly increasing number of online social network (OSN) users, study of OSN application-specific mobile network traffic has attracted a lot of research efforts in recent years. In this work, we study the temporal characteristics of Twitter traffic and propose a Twitter traffic prediction framework which combines statistical analytics and machine learning techniques. Experimental results based on real-world Twitter traffic dataset collected in the central London have shown that the proposed framework has a high prediction accuracy with low computation complexity and low demand for the size of dataset.