UCCL

Aug 2025 – Present

  • Contributed to the development of UCCL, a next-generation GPU communication library for ML workloads
  • Participated in building high-performance communication components including:
    • Collective communication (Re-architecting NCCL)
    • P2P data transfer (e.g., KV cache / weight transfer)
  • Contributed to improving performance and scalability

Heterogeneous Memory Communication

Dec 2024 – Present

  • Design and implement mechanisms for direct data transfer and unified memory management across heterogeneous GPUs (NVIDIA, AMD, Hygon, Cambricon, etc.)
  • Support distributed vLLM inference on heterogeneous GPU clusters

Kmesh

Jun 2024 – Nov 2024

Kmesh focuses on locality-aware load balancing to improve performance and reliability in distributed systems.

  • Reduce latency by routing traffic to the nearest service instances
  • Improve system availability and minimize cross-region communication overhead
  • Support fine-grained locality policies across multiple scopes:
    • region, zone, subZone, nodeName, clusterId

SATCC-framework

Oct 2022 – May 2024

  • A research framework for learning-based congestion control algorithms
  • Supports experimentation on both Linux kernel TCP and XQUIC

oncn-bwm

Jun 2023 – Oct 2023

  • Implemented ingress network QoS capabilities for mixed-workload scenarios
  • Built upon the oncn-bwm bandwidth management framework

tcp_ql

Dec 2021 – Jul 2022

  • Designed a congestion control algorithm for dynamic satellite networks
  • Implemented a Q-learning-based TCP congestion control algorithm in the Linux kernel

rl-env-mininet

Oct 2022 – May 2024

  • A research environment for ML/RL-based congestion control algorithms
  • Enables distributed reinforcement learning training on real network protocols
  • Planned for open-source release after further refinement

xquic-forrlcc

Nov 2022 – May 2024

  • Extended XQUIC to support reinforcement learning-based congestion control research

doushen

Jan 2023 – Feb 2023

  • A simplified TikTok backend built with Hertz and Kitex
  • Developed during ByteDance Youth Training Camp
  • 🏆 Awarded Best Code Quality

Multi-source Image Fusion Based on Contourlet Transform

Oct 2018 – May 2019

  • Studied multi-source image fusion methods based on the Contourlet transform
  • Focused on image processing and multi-scale geometric analysis techniques