CV

Contact

• Email: d202381502@hust.edu.cn
• Affiliations:
  
  • School of Computer Science and Technology, Huazhong University of Science and Technology
    
  • Key Laboratory of Information Storage System, Ministry of Education

Education

• 2023–present: Ph.D. student in Computer Architecture, Huazhong University of Science and Technology
• 2019–2023: B.S. in Computer Science and Technology, Beijing Jiaotong University

Research interests

• Memory systems for long-context LLM inference
• KV-cache management, offloading, compression, and scheduling
• Storage systems and heterogeneous IO paths
• CXL and memory disaggregation
• GPU memory management, GPU interconnects, and MLSys runtime support

Research experience

• Long-context LLM memory systems (2025–present)
  Advisors: Assoc. Prof. Jianxi Chen and Prof. Dan Feng, HUST
  • Study KV-cache as a managed memory object for long-context LLM inference, including tiered placement, offloading, prefetch, and scheduling.
  • Built research systems around GPU HBM, host DRAM, SSD-backed storage paths, and cross-GPU KV movement.
  • Explored exact placement, lossy eviction, calibrated saliency prediction, SLO-aware scheduling, and interconnect-aware KV handoff as separate design modes.
• LLM inference acceleration with PCIe 5.0 SSD arrays (Apr 2025–present)
  Advisors: Assoc. Prof. Jianxi Chen and Prof. Dan Feng, HUST
  • Designed a hot/cold separation scheme using attention EMA, temporal decay, and access frequency for online KV-block classification.
  • Integrated heterogeneous storage ideas from OrchFS to build a GPU–DRAM–SSD hierarchy for KV-cache management.
  • Studied overlapping compute, prefetch, and preload pipelines to reduce storage migration latency.
• GPUDirect Storage for GPU–SSD collaborative systems (Mar 2025–present)
  Advisors: Assoc. Prof. Jianxi Chen and Prof. Dan Feng, HUST
  • Explored GPUDirect Storage (GDS) for direct GPU–SSD data paths, bypassing unnecessary CPU-side copies.
  • Analyzed GPU–SSD bottlenecks and storage/runtime co-design opportunities for LLM serving.
• Stock prediction with traditional and non-traditional financial data (Jul 2022–Oct 2022)
  Advisor: Researcher Guangnan Ye, Fudan University
  • Built prediction models with TOPNet, LSTM, and LightGBM; used NLP and graph models for financial text and user-behavior signals.
  • Studied federated learning for cross-institution privacy-preserving settings and compared FedAvg and FedProx under non-IID conditions.

Work experience

• Software engineering intern — Data Platform, OLAP engine team (Nov 2023–Feb 2024)
  • Xiaomi Group
    Mentor: Yaodong Zhang, Senior Software Engineer; Apache Spark committer
  • Implemented column-level encryption for Parquet files in Apache Spark and Trino, so sensitive columns in internal OLAP pipelines could be protected independently.

Selected systems

• OrchKvCache: tiered KV-cache management across GPU HBM, host DRAM, and SSD-backed storage paths.
• HALO: identity-preserving KV tiering for memory-bounded long-context inference.
• SEER: SLO-aware KV-cache eviction and prefetch under P99 / P99.9 latency targets.
• XQP: measurement-driven audit of KV-block saliency prediction.
• PeerKV: cross-GPU KV-cache handoff contention on NVLink.

Skills

• Programming: C/C++, Java, Python, Scala, MATLAB
• Systems & tools: Git, GDB, SPDK, GPUDirect Storage, HDFS, Linux systems programming
• Research areas: storage systems, LLM inference systems, GPU memory systems, GPU interconnects, memory disaggregation, MLSys

Awards

• 2023–2025: First-class Academic Scholarship, HUST
• 2020–2021: Merit Student and Outstanding Academic Scholarship, Beijing Jiaotong University
• Nov 2020: China Undergraduate Mathematical Contest in Modeling (CUMCM), second prize

Publication record

A concise publication and manuscript list is maintained on the Publications page to avoid duplicating the same records in multiple places.