Department of Computer Science The University of Hong Kong (HKU) Room 421, Chow Yei Ching Building, HKU, Pok Fu Lam Road, Hong Kong [ Resume | Publication | Research Lab | Admitting New Students ]

I am an associate professor in HKU CS. I got my bachelor and master degrees from Tsinghua University; I joined HKU in January 2015 right after I got my PhD degree from Columbia University (my PhD supervisor is Prof. Junfeng Yang). I lead the HKU Systems Software Lab. I am interested in building parallel and distributed systems, including blockchain systems, distributed AI training/serving systems, distributed big-data and parallel computing systems, and cloud computing systems. I have a particular focus on improving the reliability, security, and performance of these systems. I publish papers in broad areas of systems, security, and networking, including SOSP, NSDI, ASPLOS, ATC, EuroSys, DSN, TPDS, and TDSC. In recent three years, I serve on the program committees of international systems/networking conferences, including OSDI, NSDI, ATC, DSN, EuroSys, SOCC, and ICDCS. I also serve as constant reviewers for international systems/networking/software/security journals, including TPDS, TOCS, TSE, TON, TMC, and TDSC. I receive several world-wide competative research awards, including a Croucher Innovation Award in 2016 (HK $5 million), an outstanding (best) paper award from ACSAC '17, two Huawei flagship research grants in 2018 (blockchain and security) and 2021 (AI), and the Best Collaborating Scientist Medal from the Huawei Theory Lab in 2021, and the RGC Research Impact Fund (RIF) in 2023 (HK $4.3 million). He is the project leader of a National Key R&D Program of China 2030 (the project topic is about building new parallel training systems for large AI models). As (one of) the project leaders or principle investigators, Dr. Cui's total amount of competative research grants in Hong Kong and mainland China has reached about HK $90 million.

My recent research papers have led to commercial software releases with global leading IT industries. For instance, My secure system papers (e.g., [Uranus AsiaCCS 2020] and [DAENet TDSC 2021]) on Trusted Execution Environments have become a core component of Huawei's Trusted and Intelligent Cloud Services (see the UTEE component in TICS). In addition, my students and I are actively collaborating with industries to jointly publish research papers and to transfer the resultant systems from these papers into commercial software of broad areas, including distributed AI training systems, permissioned blockchain systems, security and privacy preserving systems, and geo distributed transaction systems.

I admit several PhD students every year. I expect my students to have good skills/experience on hacking systems software (e.g., Linux kernel, LLVM, or distributed protocols) or AI frameworks, and have strong motivation on research. If you are interested, please directly apply here and select "systems and networking research" as your interested field during the application. If you also want to talk with me individually, please read my recent papers (at least several times for each paper), understand how they work deeply, compile and run them, and then email me what new research topics you can think of (e.g., new applications or significant improvments of my systems, or some other relevant and crazy ideas). I will reply your email quickly if your ideas make sense. I also recruit postdoc of broad systems and networking areas. Please read my papers, form a few short research proposals (ideas/plans) within the intersections of your work and my work, and send me your CV with the proposals.

I have several well funded research grants that can support student Research Assistants (RAs) and summer research interns for students around the world. If you are interested and you can work full-time in HKU for a few months, you can send me emails with your CV and thoughts on my papers.

Selected Publication ("*" means corresponding author)

• BIDL: A High-throughput, Low-latency Permissioned Blockchain Framework for Datacenter Networks [pdf | video | code]
  Ji Qi, Xusheng Chen, Yunpeng Jiang, Jianyu Jiang, Tianxiang Shen, Shixiong Zhao, Sen Wang, Gong Zhang, Li Chen, Man Ho Au, Heming Cui*
  Proceedings of the 28th ACM Symposium on Operating Systems Principles 2021 (SOSP '21). ACM results reproduced badge.
  

  Describe BIDL, a high-performance and BFT-compatible permissioned blockchain workflow that can parallelize the executions of smart contracts and their BFT ordering on dedicated datacenter networks; BIDL can also tackle aborts caused by concurrently contending smart contract transactions and non-determinism (e.g., multithreading races in transactions).

• Fold3D: Rethinking and Parallelizing Computational and Communicational Tasks in the Training of Large DNN Models [pdf | slides | code]
  Fanxin Li, Shixiong Zhao*, Yuhao Qing, Xusheng Chen, Xiuxian Guan, Sen Wang, Gong Zhang, and Heming Cui*
  IEEE Transactions on Parallel and Distributed Systems 2021 (TPDS '23)
  

  Describe Fold3D, a general pipepline parallel training system for big AI models (e.g., Transformer and GPT). Fold3D introduces a new scheduling algorithm that folds Data Parallel (DP), Pipeline Parallel (PP) and Tensor Parallel (TP) tasks in a highly parallel way. Compared to world's latest relevant training systems (e.g., Megatron), Fold3D improves all GPUs' utilization (training throughput) for training each big AI model significantly by about 40%.

• JITfuzz: Coverage-guided Fuzzing for JVM Just-in-Time Compilers
  Mingyuan Wu, Minghai Lu, Heming Cui, Yanwei Huang, Junjie Chen, Yuqun Zhang, and Lingming Zhang
  Proceedings of the 45th International Conference on Software Engineering (ICSE '23)

• CRONUS: Fault-isolated, Secure and High-performance Heterogeneous Computing for Trusted Execution Environments [pdf | slides | video | code]
  Jianyu Jiang, Qi Ji, Tianxiang Shen, Xusheng Chen, Shixiong Zhao, Sen Wang, Li Chen, Gong Zhang, Xiapu Luo, Heming Cui*
  Proceedings of the 55th ACM/IEEE International Symposium on Microarchitecture (MICRO '22). ACM results reproduced badge.
  

• ROG: A High Performance and Robust Distributed Training System for Robotic IoT [pdf | slides | video | code]
  Xiuxian Guan, Zekai Sun, Shengliang Deng, Xusheng Chen, Shixiong Zhao*, Zongyuan Zhang, Tianyang Duan, Yuexian Wang, Chenshu Wu, Yong Cui, Libo Zhang, Yanjun Wu, Rui Wang, Heming Cui
  Proceedings of the 55th ACM/IEEE International Symposium on Microarchitecture (MICRO '22). ACM results reproduced badge.
  

• A Geography-Based P2P Overlay Network for Fast and Robust Blockchain Systems [paper]
  Haoran Qiu, Tao Ji, Shixiong Zhao*, Xusheng Chen*, Ji Qi, Heming Cui, Sen Wang
  IEEE Transactions on Services Computing 2022 (TSC '22)
  

• SOTER: Guarding Black-box Inference for General Neural Networks at the Edge [pdf | slides | code]
  Tianxiang Shen, Ji Qi, Jianyu Jiang*, Xian Wang, Siyuan Wen, Xusheng Chen, Shixiong Zhao, Sen Wang, Li Chen, Xiapu Luo, Fengwei Zhang, Heming Cui
  Proceedings of the 2022 USENIX Annual Technical Conference (ATC '22)
  

  Describe SOTER, a system for enforcing the confidentiality and integrity of a trained DNN model's parameters for inference tasks on end devices by automatically splitting, transforming, recovering a part of the model's neuron layers within and outside a trusted execution environment (e.g., an Intel SGX CPU). SOTER incurs no inference precision loss for an DNN model and incurs only moderate performance overhead compared to the execution time of insecure inference tasks.

• NASPipe: High Performance and Reproducible Pipeline Parallel Supernet Training via Causal Synchronous Parallel [pdf | video | code]
  Shixiong Zhao, Fanxin Li, Xusheng Chen, Tianxiang Shen, Li Chen Sen Wang, Gong Zhang, Cheng Li, Heming Cui*
  The 2022 Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS '22). ACM results reproduced badge.
  

  Describe NASPipe, the first pipeline parallel (PP) training system for training multiple subnets concurrently in NAS (Neural Arch Search) based big AI models.

• vPipe: A Virtualized Acceleration System for Achieving Efficient and Scalable Pipeline Parallel DNN Training [pdf | code]
  Shixiong Zhao, Fanxin Li, Xusheng Chen, Xiuxian Guan, Jianyu Jiang, Dong Huang, Yuhao Qing, Sen Wang, Peng Wang, Gong Zhang, Cheng Li, Ping Luo, Heming Cui*
  IEEE Transactions on Parallel and Distributed Systems 2021 (TPDS '21)
  

  Describe vPipe, a general pipepline parallel training system for big AI models (e.g., Transformer, AboebaNet, GPT, and GNMT). vPipe tackles a severe load imbalance problem among the multi-GPU pipeline stages caused by inherent tensors accumulation in former stages, which can greatly downgrade training performance. vPipe's throughput and effective GPU ALU utilization are higher than existing premier pipepline parallel training systems (e.g., GPipe and Pipedream). vPipe is also the first pipeline parallel training system that supports NAS (Neural Arch Search) during the training of a big AI model.

• COORP: Satisfying Low-Latency and High-Throughput Requirements of Wireless Network for Coordinated Robotic Learning [pdf | code]
  Shengliang Deng, Xiuxian Guan, Zekai Sun, Shixiong Zhao, Tianxiang Shen, Xusheng Chen, Tianyang Duan, Yuexuan Wang, Jia Pan, Yanjun Wu, Libo Zhang, Heming Cui*
  IEEE Internet of Things Journal 2022 (IOT-J '22)
  

  Describe Coorp, a wireless and distributed scheduling system for robots which conduct high-throughput distributed online learning as well as low-latency wireless coordinating operations.

• Evaluating and Improving Neural Program-Smoothing-based Fuzzing [pdf]
  Mingyuan Wu, Jing Liang, Jiahong Xiang, Yuqun Zhang, Guowei Yang, Huixin Ma, Sen Nie, Shi Wu, Heming Cui, Lingming Zhang
  Proceedings of the 44th International Conference on Software Engineering (ICSE '22)
  

  Evaluate the coverage limitations of existing program smoothing fuzzing techniques and introduce our optimizations on these techniques.

• One Fuzzing Strategy to Rule Them All [pdf]
  Mingyuan Wu, Jing Liang, Jiahong Xiang, Yanwei Huang, Heming Cui, Lingming Zhang, Yuqun Zhang
  Proceedings of the 44th International Conference on Software Engineering (ICSE '22)
  

  We study the coverage limitations of existing Havoc based techniques; inspired by these findings, we further propose HavocDMA, which models the Havoc mutation strategy as a multi-armed bandit problem to be solved by dynamically adjusting the mutation strategy; HavocDMA improves the coverage of state-of-the-art.

• Achieving Low Tail-latency and High Scalability for Serializable Transactions in Edge Computing [pdf | video | code]
  Xusheng Chen, Haoze Song, Jianyu Jiang, Chaoyi Ruan, Cheng Li, Sen Wang, Gong Zhang, Reynold Cheng, Heming Cui*
  Proceedings of the European Conference on Computer Systems 2021 (EuroSys '21). ACM results reproduced badge.
  

  Describe DAST (Decentralized Anticipating and STretching), a new geo distributed transaction system for achieving serializable and low-tail latency transactions among many edge computing datacenters and regions (e.g., cities/towns).

• Efficient and DoS-resistant Consensus for Permissioned Blockchains [pdf | code]
  Xusheng Chen, Shixiong Zhao, Ji Qi, Jianyu Jiang, Haoze Song, Cheng Wang, Tsz On Li, Hubert Chan, Fengwei Zhang, Xiapu Luo, Sen Wang, Gong Zhang, Heming Cui*
  Proceedings of the 39th International Symposium on Computer Performance, Modeling, Measurements and Evaluation 2021 (Performance '21)
  

  Describe Eges, an SGX-powered distriburted consensus protocol for permissioned blockchains. Eges's new stealth consensus committee abstraction can rotate and reach consensus securely, stealthly and efficiently. Among consensus protocols for permissioned blockchains, Eges is the first protocol that can ensure both safety (consistency on confirming blocks) and good performance against tough DOS attacks, including targetting attacks on consensus nodes and network partition attacks.

• Securing Big Data Scientific Workflows via Trusted Heterogeneous Environments [paper]
  Saeid Mofrad, Ishtiaq Ahmed, Fengwei Zhang, Shiyong Lu, Ping Yang, Heming Cui
  IEEE Transactions on Dependable and Secure Computing 2022 (TDSC '22)
  

  Describe SecDataView, a secure big-data workflow exeuction system that can leverage diverse trusted execution environments on clouds.

• DAENet: Making Strong Anonymity Scale in a Fully Decentralized Network [paper | code]
  Tianxiang Shen, Jianyu Jiang, Yunpeng Jiang, Xusheng Chen, Ji Qi, Shixiong Zhao, Fengwei Zhang*, Xiapu Luo, Heming Cui*
  IEEE Transactions on Dependable and Secure Computing 2021 (TDSC '21)
  

  Describe DAENet, an SGX-powered secure communication system that can provide strong anonymity (e.g., hiding communication endpoints and traffic routes) and ensure reasonable communication efficiency even on packet dropping and DOS attacking scenarios. DAENet's strong anonimity and efficiency are scalable on Internet.

• vSMT-IO: Improving I/O Performance and Efficiency on SMT Processors in Virtualized Clouds [link]
  Weiwei Jia, Jianchen Shan, Tsz On Li, Xiaowei Shang, Heming Cui, Xiaoning Ding
  Proceedings of the 2020 USENIX Annual Technical Conference (ATC '20)
  

  Describe vSMT-IO, a virtual machine IO management system that can mitigate the IO contention in SMT (hyper-threading) in modern CPUs and greatly improve IO performance of VMs.

• HAMS: High Availability for Distributed Machine Learning Service Graphs [pdf | video | code]
  Shixiong Zhao, Xusheng Chen, Cheng Wang, Fanxin Li, Ji Qi, Heming Cui*, Cheng Li, Sen Wang
  Proceedings of the 50th IEEE/IFIP International Conference on Dependable Systems and Networks (DSN '20)
  

  Describe HAMS, an efficient fault-tolerence system for a distributed deployment graph consisting of machine learning serving models (some are stateful) with little performance overhead. HAMS is the first system that can correctly replicate the nondeterminism caused by GPU threads on running AI serving models.

• UPA: An Automated, Accurate and Efficient Differentially Private Big-data Mining System [pdf | video | code]
  Tzs On Li, Jianyu Jiang, Ji Qi, Chi Chiu So, Jiacheng Ma, Xusheng Chen, Tianxiang Shen, Heming Cui*, Yuexuan Wang, Peng Wang
  Proceedings of the 50th IEEE/IFIP International Conference on Dependable Systems and Networks (DSN '20)
  

  Describe UPA, the first automated, efficient and precise differentially private big-data mining system that can preserve individual privacy while supporting general Spark big-data queries.

• Uranus: Simple, Efficient SGX Programming and Its Applications [pdf | code]
  Jianyu Jiang, Xusheng Chen, Tzs On Li, Cheng Wang, Tianxiang Shen, Shixiong Zhao, Heming Cui*, Cho-Li Wang, Fengwei Zhang
  Proceedings of the 15th ACM ASIA Conference on Computer and Communications Security (ASIACCS '20)
  

  Describe Uranus, the first SGX-compatible and Spark-compatible secure big-data computing system that can practically support typical big-data datasets with our new big-data aware Java memory reuse protocols.

• Fulva: Efficient Live Migration for In-memory Key-Value Stores with Zero Downtime [pdf]
  Jiewen Hai, Cheng Wang, Xusheng Chen, Tsz On Li, Heming Cui*
  Proceedings of the 38th International Symposium on Reliable Distributed Systems (SRDS '19)
  

  Describe Fulva, an in-memory key-value store migration system that can almost pertain the store's normal-case performance with no downtime during migration.

• NFVactor: A Resilient NFV System using the Distributed Actor Model [pdf]
  Jingpu Duan, Xiaodong Yi, Shixiong Zhao, Chuan Wu, Heming Cui, Franck Le
  IEEE Journal on Selected Areas in Communications (JSAC) 2019
  

  Describe NFVactor, an actor-based system for efficiently and reliably deploying NFV service chains.

• Effectively Mitigating I/O Inactivity in vCPU Scheduling [pdf | code]
  Weiwei Jia, Cheng Wang, Xusheng Chen, Jianchen Shan, Xiaowei Shang, Heming Cui*, Xiaoning Ding, Luwei Cheng, F.C.M. Lau, Yuexuan Wang, Yuangang Wang
  Proceedings of the 2018 USENIX Annual Technical Conference (ATC '18)
  

  Describes vMigrator, a lightweight, easy to use tool that can effectively mitigate I/O inactivity in vCPU scheduling, greatly improving I/O performance in VMs.

• PLOVER: Fast, Multi-core Scalable Virtual Machine Fault-tolerance [pdf | video | code]
  Cheng Wang, Xusheng Chen, Weiwei Jia, Haoran Qiu, Boxuan Li, Shixiong Zhao, Heming Cui*
  Proceedings of the 15th USENIX Symposium on Networked Systems Design and Implementation 2018 (NSDI '18)
  

  Describes PLOVER, a multi-core scalable VM fault-tolerance system built on the Virtualized State Machine Replication (VSMR) concept.

• OWL: Understanding and Detecting Concurrency Attacks [pdf | code]
  Shixiong Zhao, Rui Gu, Haoran Qiu, Tsz On Li, Yuexuan Wang, Heming Cui*, Junfeng Yang
  Proceedings of The 48th IEEE/IFIP International Conference on Dependable Systems and Networks 2018 (DSN '18)
  

  Describes OWL, the first tool for detecting multi-threading concurrency attacks in general software (e.g., Linux kernels).

• How Local Information Improves Rendezvous in Cognitive Radio Networks [pdf]
  Yongqin Fu, Yuexuan Wang, Zhaoquan Gu, Xiaolin Zheng, Tianhao Wei, Zhen Cao, Heming Cui, F.C.M. Lau
  Proceedings of the IEEE International Conference on Sensing, Communication and Networking 2018 (SECON '18)
  

  Describes optimized rendezvous algorithms for CRN.

• APUS: Fast and Scalable PAXOS on RDMA [pdf | video | code]
  Cheng Wang, Jianyu Jiang, Xusheng Chen, Ning Yi, Heming Cui*
  Proceedings of the ACM Symposium on Cloud Computing (SOCC '17), 2017
  

  Describes APUS, a fast Paxos protocol and its runtime system using fast RDMA features.

• Kakute: A Precise, Unified Information Flow Analysis System for Big-data Security [pdf | video | code]
  Jianyu Jiang, Shixiong Zhao, Danish Alsayed, Yuexuan Wang, Heming Cui*, Feng Liang, Zhaoquan Gu
  Proceedings of the Annual Computer Security Applications Conference (ACSAC '17), 2017. Best paper award!
  

  Describes Kakute, an information flow tracking system for preventing various big-data security issues (data leakage in user-defined-functions) without reporting false positives.

• Confluence: Speeding Up Iterative Distributed Operations by Key-dependency-aware Partitioning
  Feng Liang, F.C.M. Lau, Heming Cui, C.L. Wang
  IEEE Transactions on Parallel and Distributed Systems 2017 (TPDS '17)
  

  Describes Confluence, a system that can greatly reduce shuffle traffic and improve performance in big-data flatforms using a novel key-dependency-aware partitioning scheme.

• Paxos Made Transparent [pdf | code]
  Heming Cui, Rui Gu, Cheng Liu, Tianyu Chen, Junfeng Yang
  Proceedings of the 25th ACM Symposium on Operating Systems Principles (SOSP '15), 2015
  

  Describes RepBox, a state machine replication system that can transparently replicate general multithreaded programs without intervention from the developers.

• Parrot: a Practical Runtime for Deterministic, Stable, and Reliable Threads [pdf | slides | video | code]
  Heming Cui, Jiri Simsa, Yi-Hong Lin, Hao Li, Ben Blum, Xinan Xu, Junfeng Yang, Garth Gibson, Randal E. Bryant
  Proceedings of the 24th ACM Symposium on Operating Systems Principles (SOSP '13), 2013
  

  Describes PARROT, a simple and deployable system for making threads deterministic and stable. PARROT is easy to use, fast, scalable, and it improves the state-space coverage of a state-of-the-art model checker, DBUG, by many orders of magnitude. Together with the code, we also released a benchmark suite with 100+ multithreaded programs, and PARROT's complete results on these programs.

• Determinism Is Not Enough: Making Parallel Programs Reliable with Stable Multithreading
  Junfeng Yang, Heming Cui, Jingyue Wu, Yang Tang, Gang Hu
  Communications of the ACM (2014)
  

  Describes our vision of stable multithreading, a radically new approach to making multithreading reliable, and summarizes our last five years of work on designing, building, and applying stable multithreading systems.

• Verifying Systems Rules Using Rule-Directed Symbolic Execution [pdf | slides]
  Heming Cui, Gang Hu, Jingyue Wu, Junfeng Yang
  Eighteenth International Conference on Architecture Support for Programming Languages and Operating Systems (ASPLOS '13), 2013
  

  Describes WOODPECKER, a system that leverages path slicing to speed up symbolic execution. It enables users to check systems rules, and avoids checking program paths irrelevant to the rule, drastically reducing the amount of redundant work.

• Sound and Precise Analysis of Parallel Programs through Schedule Specialization [pdf]
  Jingyue Wu, Yang Tang, Gang Hu, Heming Cui, Junfeng Yang
  Proceedings of the 33rd ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI '12), 2012
  

  Describes a program analysis framework for analyzing multithreaded programs with high precision. The key idea is to statically analyze a multithreaded program toward only a small set of schedules to improve precision, then enforce these schedules at runtime for soundness.

• Efficient Deterministic Multithreading through Schedule Relaxation [pdf | slides]
  Heming Cui, Jingyue Wu, John Gallagher, Huayang Guo, Junfeng Yang
  Proceedings of 23rd ACM Symposium on Operating Systems Principles (SOSP '11), 2011
  

  Describes PEREGRINE, a system for efficiently making threads deterministic, addressing a key open challenge within the field of deterministic execution.

• Stable Deterministic Multithreading through Schedule Memoization [pdf | slides]
  Heming Cui, Jingyue Wu, Chia-che Tsai, Junfeng Yang
  Proceedings of the Ninth Symposium on Operating Systems Design and Implementation (OSDI '10), 2010
  

  Describes TERN, a system for making threads more deterministic and stable. The key idea is to memoize past schedules and reuse them when possible, much like the natural tendencies in animals and humans to repeat familiar routes to avoid possible hazards along unknown routes.

• Bypassing Races in Live Applications with Execution Filters [pdf]
  Jingyue Wu, Heming Cui, Junfeng Yang
  Proceedings of the Ninth Symposium on Operating Systems Design and Implementation (OSDI '10), 2010
  

  Describes LOOM, a live-workaround system for fixing races in live applications. LOOM is safe (live-update will not introduce new errors), fast (negligible overhead for most benchmarks), and flexible (able to fix all bugs evaluated).

Members in My HKU Systems Software Lab

Alumni

Admitting New Students

Misc