星云Clustar与港科大合作论文LiteFlow被SIGCOMM 2022录用

第36届 Conference of the Special Interest Group on Data Communication ("SIGCOMM") 将于 2022 年 8 月 22日至 26日在荷兰阿姆斯特丹市召开。SIGCOMM 是计算机网络领域最高水平学术会议之一,本次投稿共 279篇,录用 55篇 (录用率为 19.7%)。

在本次会议中,香港科大智能网络与系统实验室iSING Lab和星云Clustar,UC Berkeley合作的一篇论文被录用,题目为 LiteFlow: Towards High-performance Adaptive Neural Networks for Kernel Datapath,作者为 Junxue Zhang, Chaoliang Zeng, Hong Zhang, Shuihai Hu, Kai Chen。

自适应神经网络(Adaptive Neural Network)由于其优良的性能,并能持续地学习以适应变化的环境,已被用于优化多种网络通路功能(Datapath Function),如拥塞控制、负载均衡、流量调度等。但是,对于如何高效地在最常见的内核通路 — Linux内核网络通路(Linux Kernel Datapath)部署自适应神经网络,仍处于一个未被研究的领域。直接在用户态进行神经网络部署,并通过跨内核态、用户态的通信来使用神经网络,则会造成不可避免的性能下降问题;反之,直接在内核态部署又会遭受开发难度高,额外开销大等问题。

为了有效地在内核通路部署高性能神经网络,本文提出了 LiteFlow,一种混合式的部署框架。区别于传统自适应神经网络推理、重训练的一体化设计,LiteFlow 通过解耦模型的推理以及重训练,使得其可以在各自最合适的部分进行运算。具体来说,模型推理应该直接部署在内核态,从而快速地进行响应,有效地进行报文控制,且避免用户态/内核态通信带来的巨大开销;模型重训练应该部署在用户态,从而可以直接使用用户态完善的软件、库等设施,并可以通过批量从内核台向用户态发送重训练所需要的数据来降低通信的开销。同时,LiteFlow 也设计了用户态/内核态更新同步机制,进行用户态神经网络向内核态的更新,使得内核态的神经网络也能学习、适应变化的环境。

LiteFlow 提供了一个基于 Linux Kernel v4.15.0 的实现,可以用于优化拥塞控制、流量调度、负载均衡三个网络通路功能。实验表明,LiteFlow 针对拥塞控制有 44.4% 的Goodput提升,针对流量调度以及负载均衡,则分别有33.7% 以及 56.7%的大流完成时间提升。

近年来,陈凯教授多次在IEEE、NSDI、IJCAI等学术国际会议上发表论文,其代表作有:

l (IEEE S&P’22) Han Tian, Chaoliang Zeng, Zhenghang Ren, Di Chai, Junxue Zhang, Kai Chen, Qiang Yang, “Sphinx: Enabling Privacy-Preserving Online Learning over the Cloud”, in IEEE Symposium on Security and Privacy (Oakland), 2022.

l (NSDI’22) Chaoliang Zeng, Layong Luo, Zilong Wang, Luyang Li, Wenchen Han, Nan Chen, Lebing Wan, Lichao Liu, Zhipeng Ding, Xiongfei Geng, Tao Feng, Feng Ning, Kai Chen, Chuanxiong Guo, “Tiara: A Scalable and Efficient Hardware Acceleration Architecture for Stateful Layer-4 Load Balancing”, in USENIX Symposium on Networked Systems Design and Implementation (NSDI), 2022.

l (IJCAI-FL’20) Zhaoxiong Yang, Shuihai Hu, Kai Chen, “FPGA-Based Hardware Accelerator of Homomorphic Encryption for Efficient Federated Learning”, in International Workshop on Federated Learning for User Privacy and Data Confidentiality (in Conjunction with IJCAI), IJCAI-FL, 2020 (Best Student Paper Award).

另外,陈凯教授著有多部专著作品(包括章节),其中包括:Li Chen, Justinas Lingys, Kai Chen, Xudong Liao, “Datacenter Traffic Optimization with Deep Reinforcement Learning”, invited book chapter for “Communication Networks and Service Management in the Era of Artificial Intelligence and Machine Learning”, IEEE/Wiley, 2021. 以及Kai Chen, Ankit Singla, Atul Singh, Kishore Ramachandran, Lei Xu, Yueping Zhang, Xitao Wen, Yan Chen, “OSA: An Optical Switching Architecture for Data Center Networks with Unprecedented Flexibility”, invited book chapter for “Optical Switching in Next Generation Data Centers”, Springer, 2018.等等。

陈凯教授另有发明专利20项,其中包括有:

1. Zhitang Chen, Yanhui Geng, Hong Zhang, Kai Chen, “Coflow Identification Method and System, And Server Using Method” US Patent, Application No. US16/127,649, Pub No. US10567299B2, February 2020.

2. Kai Chen, Li Chen, Bairen Yi “Mix-flow Scheduling for Commodity Datacenters”, China patent, App No. 201610639166.7, Pub No. CN106302227B, December 2019.

3. Kai Chen, Hong Zhang, Li Chen, “Co-flow Processing in RPC Communications”, China patent, App No. 201610653570.X, Pub No. CN106453112B, November 2019.

4. Kai Chen, Hong Zhang, Li Chen, “A Method for Co-flow Scheduling”, China patent, App No.

201610653395.4, Pub No. CN106656858B, October 2019.

5. Li Chen, Kai Chen, Bairen Yi, Kai Zheng, Sayee Chakravartula, Jin Zuo, “Method for Transmitting Data Streams, and Device”, US patent, App No. US16/209,699, Pub No. US20190109787A1, April 2019.

6. Yanhui Geng, Kai Chen, Qiang Yang, “Path Planning Method and Controller”, US patent, App No. US14/980,491, Pub No. US10187291B2, January 2019.

7. Kai Chen, Li Chen, “Congestion Control Methods and Congestion Window Adjustment Methods for Data Center Networks”, China patent, App No. 201610639105.0, Pub No. CN106027407B, December 2018.

8. Kun Tan, Shuihai Hu, Binzhang Fu, Kai Chen, “Data Transmission Method, Computing and Communication Device, and Data Transmission System”, China patent (filed), App No. CN201810711997.X, June 2018.

9. Feng Yuan, Kai Chen, Hong Zhang, “The Method and Server of Transmission Data”, China patent (filed), App No. CN201710020521.7A, January 2017.

10. Yongqiang Liu, Xitao Wen, Kai Chen, Yan Chen, Yong Xia, “Central Control Unit and Virtual Machine Migration Method Used for Virtual Machine Migration”, China patent, App No. CN 201110305788, Pub No. CN103023799, June 2016.

11. Zhitang Chen, Yanhui Geng, Hong Zhang, Kai Chen, “A Method and Apparatus for Co-flow Identification”, China patent (filed), App No. 201610141226.2, March 2016.

12. Cristian Lumezanu, Guofei Jiang, Kai Chen, Yueping Zhang, Vishal Singh, “Policy-aware based Method for Deployment of Enterprise Virtual Tenant Networks”, US patent, App No. US 13/740,229, Pub No. US9106576, August 2015.

13. Yanhui Geng, Kai Chen, Qiang Yang, “A Traffic-Engineering Solution in SDN based on Data Mining”, a) China patent, Application No. 201410253610.2, 2014; b) European patent, App No. EP15775601.6, 2015; c) US patent, App No. 14980491, 2015.

14. Kai Chen, Chuanxiong Guo, Haitao Wu, “Generic and Automatic Address Configuration for Data Center Networks”, US patent, App No. US 12/967,340, Pub No. US8799438, August 2014.

15. Kai Chen, Yan Chen, Xitao Wen, Yong Xia, Yongqiang Liu, “Architecture, Components and Wavelength Allocation Methods for Optical Data Center Networks”, China patent, App No. CN 201210338781, Pub No. CN103686467, March 2014.

等。