Mengwei Xu

Assistant Professor, Doctoral Advisor
State Key Laboratory of Networking and Switching Technology Computer Science Department
Beijing University of Posts and Telecommunications (BUPT)
Office: 1107, Research Building (新科研楼), BUPT (Xituchen campus), Haidian District, Beijing, P.R.China, 100876
Email: mwx AT bupt DOT edu DOT cn
Here is my Curriculum Vitae.

I'm an assistant professor in Beijing University of Posts and Telecommunications (BUPT). I received my doctoral and bachelor degrees from Peking University in 2020 and 2015, co-advised by Prof. Gang Huang and Prof. Xuanzhe Liu. I was a visiting scholar at Purdue ECE in 2019, working with Prof. Felix Xiaozhu Lin. I interned at MSRA System group from Mar. 2015 to Mar. 2016, mentored by Dr. Yunxin Liu. I also visited MSRA from Dec. 2020 through the "star-track" program.

I always look for highly self-motivated undergraduates and graduates. Please directly send your CV to me if you find my research interesting.

Research

I work on systems software, with focus on resource-constrained platforms like edge clouds, smartphones, IoTs, and satellites. Recently, I'm pushing the vision of ubiquitous learning, meaning that machine learning (training) can happen on any devices at anytime. I'm working on several related projects with my students.

Resource-efficient AI systems, to support SOTA deep learning inference and training on smartphones, wearables, and IoTs, etc. We are maintaining a paper list for this area.
Energy-efficient edge platforms and software: edge clouds need to be carbon-friendly. A position paper on it.
Satellite Computing - We take satellites as the next-generation platform that will carry exciting applications that are unaffordable on the ground. What differentiates us from others is that we have our own satellite platform in orbit!

🔥🔥 We have built an end-to-end cross-device FL platform. Here is the full code and demo. The repo will be actively maintained and you are welcome to join us!


Selected Conference Publications

Google Scholar | dblp (* = equal contributions)

[arXiv] Device-centric Federated Analytics At Ease
Li Zhang, Junji Qiu, Shangguang Wang, Mengwei Xu
-- Good for both data owners and users!
[arXiv] Understanding and Optimizing Deep Learning Cold-Start Latency on Edge Devices
Rongjie Yi, Ting Cao, Ao Zhou, Xiao Ma, Shangguang Wang, Mengwei Xu
-- Boosting your NN on devices fast!
[arXiv] Hierarchical Federated Learning through LAN-WAN Orchestration
Jinliang Yuan, Mengwei Xu, Xiao Ma, Ao Zhou, Xuanzhe Liu, Shangguang Wang.
-- Using LAN to accelerate federated learning: why not?
[MobiCom'23] Resource-Efficient Federated Learning for Modern NLP
Dongqi Cai, Yaozong Wu, Shangguang Wang, Felix Xiaozhu Lin, Mengwei Xu
International Conference on Mobile Computing and Networking 2022. Acceptance rate = 29.4% (40/136, summer round). Conditional Accept
-- Training BERT on smartphones: we can do it!
[ACM/IEEE SEC'22] Position Paper: Renovating Edge Servers with ARM SoCs
Mengwei Xu, Li Zhang, Shangguang Wang
The Seventh ACM/IEEE Symposium on Edge Computing.
-- Making edge servers energy-efficient!
[IEEE Satellite'22] A Satellite-Born Server Design with Massive Tiny Chips Towards In-Space Computing
Mengwei Xu, Li Zhang, Hongyu Li, Ruolin Xing, Qibo Sun
IEEE International Conference on Satellite Computing.
-- Maybe, that's how in-space data centers shall look like.
[MobiCom'22] Mandheling: Mixed-Precision On-Device DNN Training with DSP Offloading
Daliang Xu*, Mengwei Xu*, Qipeng Wang, Shangguang Wang, Yun Ma, Kang Huang, Gang Huang, Xin Jin, Xuanzhe Liu
International Conference on Mobile Computing and Networking 2022. Acceptance rate = 18.4% (41/223, winter round).
[arxiv] [slides] [code]
-- Remember: SoC DSPs are so energy-efficient at intensive integer operations!
[MobiSys'22] Melon: Breaking the Memory Wall for Resource-Efficient On-Device Machine Learning
Qipeng Wang*, Mengwei Xu*, Chao Jin, Xinran Dong, Jinliang Yuan, Xin Jin, Gang Huang, Yunxin Liu, Xuanzhe Liu
International Conference on Mobile Systems, Applications, and Services 2022. Acceptance rate = 21.7% (38/175).
[code] [slides]
-- Let's train large DNNs on smartphones!
[WWW'22] A Comprehensive Benchmark of Deep Learning Libraries on Mobile Devices
Qiyang Zhang, Xiang Li, Xiangying Che, Xiao Ma, Ao Zhou, Mengwei Xu, Shangguang Wang, Yun Ma, Xuanzhe Liu
The Web Conference (WWW) 2022. Acceptance rate = 17.7% (323/1822).
[slides] [code&data]
-- There are so many DL libraries... Who's the best?
[WWW'22] Commutativity-guaranteed Docker Image Reconstruction towards Effective Layer Sharing
Sisi Li, Ao Zhou, Xiao Ma, Mengwei Xu, Shangguang Wang
The Web Conference (WWW) 2022. Acceptance rate = 17.7% (323/1822).
[slides]
[IMC'21] From Cloud to Edge: A First Look at Public Edge Platforms
Mengwei Xu, Zhe Fu, Xiao Ma, Li Zhang, Yanan Li, Feng Qian, Shangguang Wang, Ke Li, Jingyu Yang, Xuanzhe Liu.
The Internet Measurement Conference (IMC). Acceptance rate = 28.1% (55/196).
[arxiv] [slides] [data]
-- Edge computing is coming!
[FSE'21] TaintStream: Fine-Grained Taint Tracking for Big Data Platforms through Dynamic Code Translation
Chengxu Yang, Yuanchun Li, Mengwei Xu, Zhenpeng Chen, Yunxin Liu, Gang Huang, Xuanzhe Liu.
The ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE) 2021. Acceptance rate = 24.5% (97/396).
[slides] [code]
[USENIX ATC'21] Video Analytics with Zero-Streaming Cameras
Mengwei Xu*, Tiantu Xu*, Yunxin Liu, and Felix Xiaozhu Lin.
USENIX ATC 2021. Acceptance rate = 18.8% (64/341). Accepted to appear
[arXiv] [MobiCom'20 Demo] [slides] [1-min Demo] [5-min Demo]
-- A totally new paradigm for video analytics systems
[EMDL'21] Towards Ubiquitous Learning: A First Measurement of On-Device Training Performance
Dongqi Cai, Qipeng Wang, Yuanqiang Liu, Yunxin Liu, Shangguang Wang, Mengwei Xu.
The International Workshop on Embedded and Mobile Deep Learning (EMDL), co-located with ACM International Conference on Mobile Systems, Applications, and Services (MobiSys) 2021.
[slides] [code]
- Time to look at the on-device training system: how can we go further?
[MM'21] Boosting Mobile CNN Inference through Semantic Memory
Yun Li, Chen Zhang, Shihao Han, Li Lyna Zhang, Baoqun Yin, Yunxin Liu, Mengwei Xu.
The ACM Multimedia (MM) 2021. Acceptance rate = 27.9% (542/1942).
[slides]
-- Learn how cache works from our brains.
[WWW'21] Characterizing Impacts of Heterogeneity in Federated Learning upon Large-Scale Smartphone Data
Chengxu Yang, Qipeng Wang, Mengwei Xu, Shangguang Wang, Kaigui Bian, Xuanzhe Liu.
The Web Conference 2021. Acceptance rate = 20.6% (357/1736).
[arXiv] [slides] [code & data]
-- Keep in mind: federated learning happens on heterogeneous smartphones!
[MobiSys'20] Approximate Query Service on Autonomous IoT Cameras
Mengwei Xu, Xiwen Zhang, Yunxin Liu, Gang Huang, Xuanzhe Liu, Felix Xiaozhu Lin.
In Proceedings of the International Conference on Mobile Systems, Applications and Services. Acceptance rate = 19.4% (34/175).
[website] [arXiv] [slides] [presentation] [short video]
-- Be intelligent and autonomous, my dear camera.
[WWW'19] A First Look at Deep Learning Apps on Smartphones
Mengwei Xu, Jiawei Liu, Yuanqiang Liu, Felix Xiaozhu Lin, Yunxin Liu, Xuanzhe Liu.
The Web Conference 2019. Acceptance rate = 18% (225/1247).
[arXiv] [slides] [code & data]
-- We have interesting findings about how DL is used in real-world smartphone apps!
[MobiCom'18] DeepCache: Principled Cache for Mobile Deep Vision
Mengwei Xu, Mengze Zhu, Yunxin Liu, Felix Xiaozhu Lin, and Xuanzhe Liu.
In Proceedings of the 24th Annual International Conference on Mobile Computing and Networking. Acceptance rate = 22% (42/187).
[arXiv] [slides] [code] [short video]
-- A fine-grained cache system to reduce CNN computations on redundant input.
[IMWUT'18] DeepType: On-Device Deep Learning for Input Personalization Service with Minimal Privacy Concern
Mengwei Xu, Feng Qian, Qiaozhu Mei, Kang Huang, and Xuanzhe Liu.
In Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies.
[slides]
-- AFAIK, the first work to demonstrate the benefit and feasibility to train DNNs on smartphones.
[IMWUT'18] PrivacyShield: A Mobile System for Supporting Subtle Just-in-time Privacy Provisioning through Off-Screen-based Touch Gestures
Saumay Pushp, Yunxin Liu, Mengwei Xu, Changyoung Koh, and Junehwa Song.
In Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies.
[EuroSys'18] Power Sandbox: Power Awareness Redefined
Liwei Guo*, Tiantu Xu*, Mengwei Xu, Xuanzhe Liu, and Felix Xiaozhu Lin.
In Proceedings of the 13th EuroSys Conference. Acceptance rate = 16% (43/262).
[WWW'17] AppHolmes: Detecting and Characterizing App Collusion among Third-Party Android Markets
Mengwei Xu, Yun Ma, Xuanzhe Liu, Felix Xiaozhu Lin and Yunxin Liu.
In Proceedings of the 26th International World Wide Web Conference. Acceptance rate = 17% (164/966).
-- Your apps can maliciously start each other from background.

Selected Journal Publications

[TMC'22] FLASH: Heterogeneity-Aware Federated Learning at Scale
Chengxu Yang, Mengwei Xu, , Qipeng Wang, Zhenpeng Chen, Kang Huang, Yun Ma, Kaigui Bian, Gang Huang, Yunxin Liu, Xin Jin, and Xuanzhe Liu
IEEE Transactions on Mobile Computing.
[IEEE Pervasive Computing'21] A Case for Camera-as-a-Service
Mengwei Xu, Yunxin Liu, Xuanzhe Liu.
IEEE Pervasive Computing, Special Issue on Pervasive Video and Audio.
-- Our ambitious vision about future video systems
[JoS'20] Autonomous Learning System towards Mobile Intelligence (in Chinese)
Mengwei Xu, Yuanqiang Liu, Kang Huang, Xuanzhe Liu, Gang Huang.
Journal of Software.
[TMC'19] DeepWear: Adaptive Local Offloading for On-Wearable Deep Learning
Mengwei Xu, Feng Qian, Mengze Zhu, Feifan Huang, Saumay Pushp, Xuanzhe Liu.
IEEE Transactions on Mobile Computing.
-- Your wearables deserve deep learning as well!
[TMC'17] ShuffleDog: Characterizing and Adapting User-Perceived Latency of Android App
Gang Huang, Mengwei Xu, Felix Xiaozhu Lin, Yun Ma, Yunxin Liu, Saumay Pushp, Xuanzhe Liu.
The International Workshop on Embedded and Mobile Deep Learning (EMDL), co-located with ACM International Conference on Mobile Systems, Applications, and Services (MobiSys) 2021.
IEEE Transactions on Mobile Computing.
-- Multi-task GPU scheduling on smartphones.

Academic Services

  • TPC Member
    2023: MobiSys;
    2022: ICWS, SCC, ICCCN;
    2021: ICDCS, BigCom, MobiArch@MobiCom, ICWS;
    2019: MobiSys Rising Stars Forum;
  • Program Chair
    IEEE SAGC 2022;
  • Vice Program Chair
    IEEE SAGC 2021;
  • Session Chair
    IEEE CLOUD 2021;
  • Secretary of IEEE Technical Community on Services Computing (2022-)

Teaching

  • Operating System (Fall 2022)
  • Previously in PKU, I have worked as TA for Distributed Machine Learning Systems (Fall 2019), Operating System Practice (Fall 2016, Spring 2017), Introduction to Computing (Fall 2014, 2016)

Invited Talks


Materials


Last Updated: Oct. 2022