Home - Zhuang Wang

Publications

Zhuang Wang, Ke Liu, Weiyi Chen, Mingyu Chen, Lixin Zhang. A Novel Approach for All-to-All Routing in All-optical Hypersquare Torus Network. ACM International Conference on Computing Frontiers (CF), 2016. PDF Code
Ke Liu, Zhuang Wang, Jack Y. B. Lee, Mingyu Chen, Lixin Zhang. Adapative Rate Control over Mobile Quality of Servie. ACM/IEEE IWQoS, 2016. PDF
Zhuang Wang, Ke Liu, Yifan Shen, Mingyu Chen, Lixin Zhang. Intra-host Rate Control with Centralized Approach. IEEE Cluster, 2016. PDF
Zhuang Wang, Weifa Liang, Meitian Huang, Yu Ma. Delay-Energy Joint Optimization for Task Offloading in Mobile Edge Computing. arXiv:1804.10416, 2018. PDF

Sep. 2014 - June. 2017	M. E., Computer Science Institute and Computing Technology. Chinese Academy of Sciences, Beijing, China GPA: 87.9/100, Advised by Prof. Mingyu Chen Thesis: An Application-driven Flow Scheduling in Data Centers (in Chinese)
Sept. 2010 - June. 2014	B. E., Computer Science Huazhong University of Science & Technology (HUST), Wuhan, China GPA: 87.66/100, 7/320

Awards	Organization	Date
National Scholarship	Ministry of Education of P. R. of China	2011
National Encouragement Scholarship	Ministry of Education of P. R. of China	2013
National Scholarship	Ministry of Education of P. R. of China	2016
International Postgraduate Research Scholarship	Australian government	2017

Jun. 2019 - Present	King Abdullah University of Science and Technology Advised by Prof. Marco Canini Scale machine learning with in-network aggregation. [see DAIET]
Aug. 2018 - Dec. 2018	Brown University Advised by Prof. Theophilus Benson Probe TCP flows and packets information from the kernel with eBPF Trace packet transmission across different layers in the kernel Measure NIC-to-NIC latencies Diagnose performance problems in microservices architecture
Sep. 2017 - May. 2018	Australian National University, Canberra, Australia Advised by Prof. Weifa Liang Designed task offloading algorithms in Mobile Edge Computing
Jan. 2014 - May. 2014	Institue of Computing Technology, Chinese Academy of Sciences, Beijing, China Advised by Prof. Binzhang Fu Worked on congestion prediction in Network on Chips (NoC)

μBPF, a performance diagnosis framework in microservices architecture [GitHub]
μBPF leverages eBPF to capture TCP metrics from the kernel for performance diagnosis in microservices architectures. It probes two levels of metrics: flows and packets. The flow-level metrics include sixteen elements, such as flight size, CWnd and sampled RTT. The packets-level metrics are the breakdown of RTTs, including latencies in TCP layer, IP layer, MAC layer and the network (from NIC to NIC). μBPF solved several challenges, such as network address translation in microservices architecture, clock synchronization and trace sampling.

Haiyun, a labeled network system [GitHub]
Haiyun enables NICs to distinguish packets with different service-level agreements (SLAs). With the support of DPDK and mTCP, Haiyun could significantly reduce the long tail delay in data centers by keeping packets with different SLAs in different queues from NICs to applications.
My work consists of two components.
One is a flow generator, which could generate more than 1 million concurrent TCP connections with just one physical server. The other is a third-party tool to measure the performance of Haiyun. To measure the delay resulted from packets processing in servers, we use switches’ mirror function to monitor packets and measure their delays on a specific server. To reduce the overheads of operating system, we leverage DPDK to bypass the kernel to optimize the performance of this tool.

Panda, a flow scheduling in data centers [GitHub]
Panda emulates shortest job first (SJF) scheduling to optimize the average flow completion time for throughput-intensive applications on the premise that flow information is not known a priori. Also, it bounds low latency for delay-sensitive applications by leveraging the distinct flow size distributions of the two kinds of applications. At its heart, Panda derives an optimal threshold to divide packets into two categories: large and small, ensuring that small packets dominate traffic from delay-sensitive applications and large ones dominate traffic from throughput-intensive applications. Panda then allocates each flow a counter which is initiated with zero. Large packets increase the counter while small packets decrease it. Then Panda assigns priorities to flows according to their counters.