ACM SIGCOMM 2018 Workshop on Network Meets AI & ML (NetAI 2018)

Workshop Program

Friday, August 24, 2018, InterContinental

8:30 am - 8:50 am Opening

Session Chair: Marco Canini (KAUST, KSA) and Chen Tian (NJU, China)

Location: InterContinental, Ballroom I

Opening

8:50 am - 9:40 am Keynote I: Toward Highly-Available, Intelligent Cloud and ML Systems

Speaker: Chuanxiong Guo (ByteDance, China)

Location: InterContinental, Ballroom I

Abstract: In this talk, I will try to bridge two seemly remote areas, system/networking and machine learning, together. In the first part, I will present our on-going journey of building highly-available cloud systems, by leveraging the power of machine learning algorithms for failure detection and localization. Specifically, I will focus on the Deepview project which uses machine learning techniques for VM VHD failure localization. In the second part, I will present how a renewed networking technique, RDMA, can be used to scale and accelerate machine learning training. Through these two projects, I hope to demonstrate that the marriage of system/networking and machine learning are much desired and inevitable. We need machine learning to bring more intelligence into large-scale systems and to further improve system availability, and we need the principles and techniques from system and networking to accelerate machine learning algorithms at scale. Through the interactions of system/networking and machine learning, more intelligent and highly available cloud and ML systems can be built in the near future.

Bio: Chuanxiong Guo is a Director of the AI Lab of Bytedance. Before that, he was a Principal Researcher at Microsoft Research Redmond. He spent several years in Microsoft Azure as a Principal Software Engineering Manager. He was a Senior Researcher at Microsoft Research Asia before he joined Microsoft Azure. His areas of interest include networked systems at large-scale, data center networking, cloud computing, systems availability and troubleshooting. He is among the first to start data center networking research. He has won several best paper awards for his research in data center networking. Several of his envisions including DCN virtualization, DCN monitoring, and ServerSwitch generated both academic and industrial impacts. Many of the systems that he has designed and implemented, including RDMA/RoCEv2 and Pingmesh are deployed in Microsoft production data centers, are running in millions of servers in data centers, and widely adopted by the cloud computing industry.

9:40 am - 10:30 am Session I: Systems

Location: InterContinental, Ballroom I

Session I: Systems

9:40 am - 10:05 am

HiPS: Hierarchical Parameter Synchronization in Large-Scale Distributed Machine Learning

Jinkun Geng, Dan Li, Yang Cheng, Shuai Wang, Junfeng Li (Tsinghua, China)

10:05 am - 10:30 am

DeepConfig: Automating Data Center Network Topologies Management with Machine Learning

Saim Salman (Brown, USA), Chris Streiffer (Duke, USA), Huan Chen (UESTC, China), Theophilus Benson (Brown, USA), Asim Kadav (NEC, USA)

10:30 am - 11:00 am Tea/Coffee Break

Location: InterContinental, Pre-Function Area
Tea/Coffee Break

11:00 am - 11:50 am Keynote II: Self-Programming Networks: Architecture and Algorithms

Speaker: Balaji Prabhakar (Stanford, USA)

Location: InterContinental, Ballroom I

Abstract: In this talk, we describe Self-Programming Networks (SPNs), an ongoing research effort at Stanford for making cloud computing networks autonomous; that is, to enable networks to sense and monitor themselves, and program and control themselves. We describe how AI/ML techniques can be used to achieve this goal. We present the architecture of SPNs and key algorithms: (i) Simon, for fine-grained network measurement using packet and probe timestamps taken at the network’s edge, and (ii) Huygens, for nanosecond-level clock synchronization in real-time and at scale. We will focus on Huygens and present results from some deployments.

Bio: Balaji Prabhakar is VMWare Founders Professor of Computer Science at Stanford University, a faculty member in the Electrical Engineering and Computer Science departments. He’s interested in cloud computing and data center networks, and leads the Self-Programming Networks research group at Stanford with Mendel Rosenblum. During 2007—2010 he co-invented the QCN (Quantized Congestion Notification) algorithm, and led the effort to standardize it as the IEEE 802.1Qau Data Center Ethernet Congestion Notification standard. QCN’s non-binary signaling of congestion inspired the widely-used DCTCP algorithm and the DCQCN algorithm. From 2008 he has worked on Societal Networks: networks vital for society’s functioning, such as transportation, electricity and recycling systems. He worked on developing and deploying "nudge engines" for transportation systems (notably, Singapore mass transit and the Bay Area Rapid Transit (BART)), wellness programs, and corporate learning programs. Based on this work, he co-founded Urban Engines, a startup which was acquired by Google in 2016. Balaji has received the NSF CAREER award, the Erlang Prize from the INFORMS Applied Probability Society, the Rollo Davidson Prize from the University of Cambridge given to "young scientists of outstanding promise and achievements in probability, statistics and related fields," and delivered the Lunteren Lectures of the Dutch Operations Research Society. He is a Fellow of IEEE, ACM and the Alfred P. Sloan Foundation. He is the recipient of the inaugural IEEE Innovation in Societal Infrastructure Award which recognizes "significant technological achievements and contributions to the establishment, development and proliferation of innovative societal infrastructure systems.” He is a co-recipient of best paper awards from IEEE Infocom, IEEE HotNets, and ACM SIGMETRICS; and a “test of time award” from ACM Sigmetrics. He has served on the Advisory Board of the Future Urban Mobility Initiative of the World Economic Forum,and served on a Panel of Judges in Governor’s Andew Cuomo’s "MTA Genius Challenge” for addressing the problems faced by New York City’s Subway system.

11:50 am - 12:40 pm Session II: Measurements I

Location: InterContinental, Ballroom I

Session II: Measurements I

11:50 am - 12:15 pm

Empowering Sketches with Machine Learning for Network Measurements

Tong Yang, Lun Wang (PKU, China), Yulong Shen (Xidian, China), Muhammad Shahzad (NCSU, USA), Qun Huang (CAS, China), Xiaohong Jiang (FUN, Japan), Kun Tan (Huawei, China), Xiaoming Li (PKU, China)

12:15 pm - 12:40 pm

Efficient Distribution-Derived Features for High-Speed Encrypted Flow Classification

Johan Garcia, Topi Korhonen (KAU, Sweden)

12:40 pm - 2:00 pm Lunch Break

Location: InterContinental, Pre-Function Area
Lunch Break

2:00 pm - 3:15 pm Session III: Mobility

Location: InterContinental, Ballroom I

Session III: Mobility

2:00 pm - 2:25 pm

Adaptive Multiple Non-negative Matrix Factorization for Temporal Link Prediction in Dynamic Networks

Kai Lei, Meng Qin (PKU, China), Bo Bai, Gong Zhang (Huawei, China)

2:25 pm - 2:50 pm

Tracking Groups in Mobile Network Traces

Kun Tu (UMass, USA), Bruno Ribeiro (Purdue, USA), Ananthram Swami (ARL, USA), Don Towsley (UMass, USA)

2:50 pm - 3:15 pm

Wireless Optimisation via Convex Bandits: Unlicensed LTE/WiFi Coexistence

Cristina Cano (UOC, Spain), Gergely Neu (UPF, Spain)

3:15 pm - 3:45 pm Tea/Coffee Break

Location: InterContinental, Pre-Function Area
Tea/Coffee Break

3:45 pm - 5:00 pm Session IV: Caching, Forwarding and TCP

Location: InterContinental, Ballroom I

Session IV: Caching, Forwarding and TCP

3:45 pm - 4:10 pm

DeepCache: A Deep Learning Based Framework For Content Caching

Arvind Narayanan, Saurabh Verma, Eman Ramadan, Pariya Babaie, Zhi-Li Zhang (UMN, USA)

4:10 pm - 4:35 pm

IFS-RL: An Intelligent Forwarding Strategy Based on Reinforcement Learning in Named-Data Networking

Yi Zhang (PKU, China), Bo Bai (Huawei, China), Kuai Xu (ASU, USA), Kai Lei (PKU, China)

4:35 pm - 5:00 pm

Improving TCP Congestion Control with Machine Intelligence

Yiming Kong (Georgia Tech, USA), Hui Zang (FutureWei, USA), Xiaoli Ma (Georgia Tech, USA)

5:00 pm - 5:05 pm Mini-Break

Location: InterContinental, Pre-Function Area
Mini-Break

5:05 pm - 5:55 pm Session V: Measurements II

Location: InterContinental, Ballroom I

Session V: Measurements II

5:05 pm - 5:30 pm

Deep-Q: Traffic-driven QoS Inference using Deep Generative Network

Shihan Xiao, Dongdong He, Zhibo Gong (Huawei, China)

5:30 pm - 5:55 pm

Assessing the Impact of Network Events with User Feedback

Shobha Venkataraman, Jia Wang (AT&T, USA)

5:55 pm - 6:00 pm Closing

Location: InterContinental, Ballroom I

Closing

Overview

Distributed processing systems for Artificial Intelligence (AI) and Machine Learning (ML), such as Hadoop, Spark, Storm, GraphLab, TensorFlow etc., are widely used by industry. On the one hand, networking is a well-known bottleneck for AI & ML systems. For example, traffic patterns experienced by ML applications during training can be optimized using techniques such as parameter servers and vertex-cut for graphs. New technologies are also being embedded in the network; examples include RDMA over converged Ethernet (RoCE) and GPU direct. On the other hand, the ever increasing complexity of networks makes effective monitoring, modeling, auditing, and overall control of network traffic difficult if not impossible. Hence there is a need for more powerful methods to solve the challenges faced in network design, deployment, and management. AI & ML have been successfully applied to various perceptual domains, including computer vision, natural language processing, and voice recognition. In addition, ML techniques are showing impressive results in new domains such as medicine, finance and astronomy, to name a few. This success in non-perceptual domains suggests that ML techniques could be successfully applied to problems in the network space. The networking community should look upon all these challenges as their opportunities in the Machine Learning era.

Call for Papers

NetAI 2018 provides a venue for presenting innovative ideas to discuss future research agendas in computer networking of/by/for AI & ML systems. We look for submissions of previously unpublished work on topics including, but not limited to, the following:

telemetry in DC
network measurement
closed loop control
AI & ML algorithm for network scheduling and control
self learning network architecture and system
services identify with AI & ML
traffic engineer with AI & ML
congestion control based on AI & ML
network security based on AI & ML
network verification
network QoS based on AI & ML
measurement and analysis of network traffic for AI & ML systems
optimizing AI & ML systems with new networking options (e.g., RDMA, NVLink)
AI & ML application-driven networking optimization (e.g., MPI offload)
networking AI & ML computation with large-scale storage technologies (e.g., NVMe over Fabric, FCoE)
networking for distributed ML systems
networking for large-scale distributed graph processing systems
new network functionality for AI & ML system ( e.g., computation in network )
application-driven network architecture design

Keynotes

Balaji Prabhakar

Stanford, USA
Chuanxiong Guo

ByteDance, China

Submission Instructions

Submission implies the willingness of at least one author to attend the workshop and present the paper. Accepted papers will be included in the ACM SIGCOMM conference proceedings. Workshop organizers reserve the right to exclude a paper from distribution after the workshop if the paper is not presented at the conference. Submitted papers should contain original material that have not be previously published, nor currently submitted for consideration elsewhere. All submissions should be written in English, with a maximum length limit of 6 printed pages, including all the figures, references and appendices. The ACM SIGCOMM LaTeX and Microsoft Word templates, as well as related information, can be found at the ACM SIG Proceedings website.

Specifically, please note the following format requirements:

All papers submitted will be peer-reviewed (single-blind) and evaluated based on their suitability (i.e., within the workshop scope), novelty, and merit.
All papers must be submitted in Portable Document Format (PDF) electronically. Whatever text processor or formatter you use to write your paper (LaTeX, Microsoft Word, FrameMaker, etc.), please convert the output to PDF before submission.
Submissions must be limited to 6 pages, single-spaced, double-column, with each column 9.25” by 3.33”, 0.33” space between columns, use at least a 10pt font, and be correctly formatted to be printed on letter-sized (8.5” by 11”) paper. Submissions that violate the formatting requirements may not be reviewed.
We must be able to print your paper once it is submitted. Therefore, all fonts used in the paper MUST be embedded in the submitted PDF file.
The paper must be in a state where it can be printed clearly on standard black-and-white printers. Reviewers are not required to view your paper in color.

Please submit your paper via https://sigcomm18netaim.hotcrp.com/

Authors Take Note

The official publication date is the date the proceedings are made available in the ACM Digital Library. This date may be up to TWO WEEKS prior to the first day of the conference. The official publication date affects the deadline for any patent filings related to published work.

Registration

Attendance of the workshop is by open registration and subject to the same registration fees and rules as all the other SIGCOMM 2018 workshops. The registrants of the workshop may freely attend any workshop on the same day.