Executing tasks with changing dependencies
SAGE organizes multi-stage tasks. MorphStream adjusts concurrent execution based on runtime dependencies.
SAGE · ICML 2026 · MorphStream · SIGMOD 2023HUST · SYSTEMS RESEARCH
Data-intensive systems in dynamic environments
Task execution · Information maintenance · Online service control
School of Computer Science and Technology · HUST
Shuhao Zhang Professor, HUST
Shuhao Zhang is a professor in the School of Computer Science and Technology at Huazhong University of Science and Technology (HUST). He leads a task under a major national science and technology project of China’s Ministry of Science and Technology. He received the NSFC Excellent Young Scientists Fund (Overseas) and was awarded support through the NSFC Research Fund for International Excellent Young Scientists (RFIS-II). Before joining HUST, he was an assistant professor at Nanyang Technological University and Singapore University of Technology and Design, and a postdoctoral researcher at Technische Universität Berlin.
His research focuses on data-intensive systems in dynamic environments. His group studies how systems can remain efficient, controllable, and reliable as task dependencies, available information, and runtime resources change. The group works on dependency-driven task execution, maintenance and reuse of evolving information, and online service control under resource constraints. It also studies how to compose mechanisms through explicit contracts and make state a first-class concern in execution.
Current work spans LLM inference, dynamic retrieval, and stateful data processing. Systems such as vLLM-HUST, ECPA, and StateAxis connect mechanism design, engineering, and evaluation on real workloads.
01 / RESEARCH
Research
Task dependencies, data, and available resources change while systems are running. We study how systems can adapt to these changes in the way they execute tasks, maintain information, and manage online services.
DATA-INTENSIVE SYSTEMS
Data-intensive systems in dynamic environments
Adapting to changes in tasks, information, and resources
Three research directions
Dependency-driven task execution
Multi-stage inference
Concurrent and heterogeneous execution
Maintenance and reuse of evolving information
Dynamic data and retrieval
Past results and external memory
Online service control under resource constraints
Changes in load and capacity
Recovery and state migration
Systems that connect the research directions
MECHANISMS
ECPA · Mechanism contracts and composition
Define how extensions declare capabilities, join the system, are validated and isolated, and are removed.
STATE
StateAxis · State-centric execution
Give execution and serving a consistent model of state ownership, layout, access, migration, and lifecycle.
Evaluation tools · Workloads, benchmarks, profiling, fault injection, and regression testing
Current focus · High-performance inference on Chinese accelerators
Use vLLM-HUST to evaluate system mechanisms with real models, hardware, and workloads.
ECPA and StateAxis connect the three research directions through contracts for mechanisms and a shared model of state. Application workloads help us test how broadly these methods apply.
02 / SYSTEMS & RESULTS
Selected Results
Maintaining and reusing evolving information
GRACE reduces the cost of maintaining dynamic graph structures. BriskSeed and Neuromem study the reuse of search history and external memory.
GRACE · ICDE 2026 · Related systems (Chinese) →Operating services under resource constraints
BidKV addresses memory pressure in LLM serving. Spacker supports state migration and resource adjustment while stream processing jobs are running.
BidKV · SC 2026 · Spacker · ICDCS 2025Systems and open-source projects (Chinese) · Full publication archive and PDFs →
03 / APPLICATIONS
Applications and Evaluation
Open systems and real workloads
We implement our methods in working systems and evaluate them on real workloads. Our current focus is vLLM-HUST, an inference engine for Chinese accelerators. We continue to share code for task execution, dynamic retrieval, state management, and evaluation.
Systems and open-source projects (Chinese)
Applications
Our faculty digital twin provides an ongoing question-answering workload for studying inference requests, memory updates, and retrieval. Other collaborations include power grid operations, inference for bioinformatics models, real-time industrial fault diagnosis, and low-latency audio and video communication. These applications help us test how our methods behave under different workloads, hardware, and reliability requirements.
Try the faculty digital twin ↗ · Paper on real-time industrial fault diagnosis
04 / PEOPLE & COLLABORATION
People and Collaboration
We welcome applications from prospective master’s and PhD students and long-term interns. Students can work on task execution, information maintenance, or online service control, and contribute to systems such as ECPA, StateAxis, and vLLM-HUST. Projects span problem analysis, mechanism design, implementation, and experiments on real hardware and workloads. We welcome backgrounds in distributed systems, operating systems, computer architecture, compilers, and performance optimization.
Applications and collaboration. Please email shuhao_zhang@hust.edu.cn with a short description of the problem you want to work on or your application needs. Students should include a CV, relevant project experience, or links to code.
HUST application portal. Select Computer Architecture (计算机系统结构) as your research direction.
05 / NEWS
News
News and paper acceptances
- November 2026. BidKV is scheduled for November 18 in the Resource-Aware LLM Serving Pipelines session at SC26. Conference program
- September 2026. BriskSeed was accepted in the first research round of ICDE 2027. Official acceptance list · Paper overview
- September 2026. Awarded support through the NSFC Research Fund for International Excellent Young Scientists (RFIS-II).
- August 2026. Our article on BidKV appeared in the ZTE Technology Journal special issue on compute and storage for LLM inference. Journal article (Chinese)
- July 2026 one paper accepted to SC 2026: BidKV
- July 2026 two system demonstrations accepted to ICPP 2026: BriskSnapshot and SAGE
- May 2026 two papers accepted to ICML 2026, including Neuromem and SAGE
- Apr 2026 two collaborative papers accepted to ACL 2026, including one Main Conference paper and one Findings paper
- Mar 2026 recruiting students interested in LLM inference engines and serving systems; prospective students are welcome to contact me by email with a CV and relevant project or code links
- Jan 2026 two papers accepted to WWW 2026
- Nov 2025 launch of our LLM inference system: https://ws.sage.org.ai/
- Nov 2025 One paper accepted to SIGMOD 2026
- Nov 2025 One paper accepted to TKDE 2026
- Nov 2025 One paper accepted to AAAI 2026
- Oct 2025 One paper accepted to ICDE 2026
- July 2025 One paper accepted to NDBC 2025
- July 2025 One paper accepted to NDBC 2025
- July 2025 One paper accepted to VLDB 2025
- March 2025 One paper accepted to ICDCS 2025
- March 2025 One paper accepted to TKDE 2025
- March 2025 One paper accepted to ICDE 2025
- March 2025 One paper accepted to CVPR 2025
06 / RESOURCES
Teaching and Resources
- Courses · Database Systems and Implementation · LLM Inference Systems and Practice · Graduate Research Paper Writing
- Profile · CV · Bio in English and Chinese · Research funding, honors, and academic service (Chinese)
- Surveys (preprints) · Inference engines for Chinese accelerators · Parallel and distributed state management