I’m a PhD student in Computer Science at North Carolina State University, working in the NCSU Systems Lab with Prof. Frank Mueller. Before NC State, I received my M.S. from the Wuhan National Laboratory for Optoelectronics (WNLO) at Huazhong University of Science and Technology (HUST), advised by Prof. Ke Zhou, and my B.Eng. from Hainan University.

My research focuses on systems for HPC and AI workloads, especially the memory and storage hierarchy:

  • Checkpointing and fault tolerance: offloading checkpoint data movement for large-scale training and HPC to SmartNICs (DPUs).
  • Deep learning I/O: multi-tier caching that serves training data from the fastest of memory, local NVMe, and parallel file systems.
  • Memory tiering: managing heterogeneous memory, including CXL memory and persistent memory.

🌎 Motto

  • 读万卷书 行万里路
  • Gain wisdom through reading and broaden horizons through travel.

🔥 News

  • 2026.11:  🎉 Passed the PhD Oral Preliminary Exam. Thanks to everyone for the support!
  • 2025.12:  🏆 Our work FitCache has been accepted to IPDPS 2026. See you in New Orleans!
  • 2025.10:  🎉 Passed the PhD Written Preliminary Exam. Thanks to everyone for the support!
  • 2024.10:  🏆 Poster accepted at SC 2024. Looking forward to presenting in Atlanta.
  • 2023.08:  🎓 Started my PhD journey at NC State University!

📖 Educations

  • 08/2023 – Present, North Carolina State University, Raleigh, USA
    Ph.D. candidate in Computer Science (Advisor: Prof. Frank Mueller)

  • 09/2019 – 06/2023, Huazhong University of Science and Technology, Wuhan, China
    M.Sc. in Computer System Architecture
    • Thesis: Research on Performance Anomaly Detection and Root Cause Analysis for Storage Systems in Data Centers
    • Advisor: Prof. Ke Zhou
    • Focus: Data Center, Storage System, Anomaly Detection, Root Cause Analysis, Machine Learning
  • 09/2015 – 06/2019, Hainan University, Haikou, China
    B.Eng. in Computer Science and Technology

📝 Publications

* These authors contributed equally to this work.

  1. Guangxing Hu, Awais Khan, Christopher zimmer, Michael Brim, Frank Mueller.
    “FitCache: A Transparent Drop-In Framework for Multi-Tier Caching to Accelerate Distributed Deep Learning Workloads.”
    IPDPS 2026, New Orleans, USA.

  2. Guangxing Hu, Awais Khan, Frank Mueller.
    “A Zero-Copy Storage with Metadata-Driven File Management Using Persistent Memory.”
    SC Poster, 2024, Atlanta, USA.

  3. Yu Liu, Yunchuan Guan, Tianming Jiang, Ke Zhou, Hua Wang, Guangxing Hu, Ji Zhang, Wei Fang, Zhuo Cheng, Ping Huang.
    “SPAE: Lifelong Disk Failure Prediction via End-to-End GAN-based Anomaly Detection with Ensemble Update.”
    Future Generation Computer Systems, 2023.

  4. Yanzhao Xie*, Guangxing Hu*, Yu Liu, Zhiqiu Lin, Ke Zhou, Yuhong Zhao.
    “How Visual Chirality Affects the Performance of Image Hashing.”
    Neural Computing and Applications, 2022.

  5. Yangtao Wang, Yanzhao Xie, Lisheng Fan, Guangxing Hu.
    “STMG: Swin transformer for multi-label image recognition with graph convolution network.”
    Neural Computing and Applications, 2021.

🎖 Honors and Awards

  • National Grand Prize of Massive Storage Algorithm Competition 2023
  • Academic Scholarship of Huazhong University of Science and Technology 2019–2022
  • Grand Prize, Group Programming Ladder Tournament, China Collegiate Computing Contest 2018
  • The Principal Scholarships of Hainan University (Top 2%) 2018

💻 Internships

  • 2025.06 - 2025.08, Oak Ridge National Laboratory (ORNL), Oak Ridge, TN, USA.
    Designed FitCache, a multi-tier caching system for accelerating deep learning training. Implemented coordinated caching across memory and storage tiers to reduce data access latency, and evaluated the system at world-leading supercomputer Frontier. This work was presented at IPDPS 2026.

  • 2024.06 - 2024.08, Oak Ridge National Laboratory (ORNL), Oak Ridge, TN, USA.
    Extended the HVAC deep learning caching framework with a Persistent Memory–based cache tier, enabling zero-copy access via DAX. Deployed and debugged the system on Frontier and ARC clusters, addressing MPI and runtime issues across environments. Achieved preliminary I/O performance improvements and presented the work in an SC 2024 poster.

💬 Talks

  • 2025.08, Leveraging Memory Hierarchies for Scalable Caching in Deep Learning I/O on HPC Systems, Oak Ridge, TN, USA.

🔖 Projects

(a) Bulk writes to DRAM cut a co-located DRAM reader from 68.6 to 10.9 GiB/s, while redirecting them to CXL restores 68.6 GiB/s; (b) a unified DRAM/CXL manager with future-value admission, an interference guard, and placement, retention, and movement decisions

Unified DRAM/CXL Memory Management

In progress · Interference-aware tiering for LLM serving and training

LLM serving and training produce bulk writes, such as KV-cache blocks and checkpoints, that can saturate DRAM write bandwidth and starve co-located readers. Instead of throttling these writers, this project treats CXL memory as an independent, writable memory pipe: bulk writes land on CXL by default, and DRAM and CXL are managed as one pool with separate capacity and write-bandwidth budgets. The prototype runs on real CXL hardware (Micron CZ120 memory expanders).

  • CXL-first write path: bulk writes go to a CXL-resident segment store; DRAM is borrowed only when a latency probe shows headroom, the loan can be revoked at no cost, and data drains to NVMe without passing through DRAM
  • Interference guard: an online, SLO-safe DRAM write budget with arbitration across concurrent writers such as checkpointing and KV-cache traffic
  • Reuse awareness: placement, retention, and movement follow each block's future value versus its movement cost; a self-labeled admission filter learns which new KV-cache blocks are not worth storing

Results so far: alongside a moderate co-located workload, the CXL-first write path is 3× faster than throttling and causes 4–10× less interference than DRAM staging; it halves PyTorch DCP async_save blocking next to an Adam optimizer (1.6 s to 0.8 s); and the admission filter beats SLRU on two real traces while cutting cache write traffic by over two-thirds.

DPUBridge overview: (a) architecture with the control layer on the host and the data layer on the DPU; (b) workflow with one-time region setup and a descriptor-based hot path

DPUBridge

Under review · DPU-driven checkpoint relay over GVMI and RDMA

DPUBridge takes the repeated checkpoint data path off the host CPU. The checkpoint framework still decides what to save and when, while the DPU reads checkpoint data straight from host memory and relays it over RDMA to a peer node. The peer places an off-node recovery copy in remote host memory and lazily flushes it to NVMe.

  • Register once, reuse every checkpoint: each checkpoint region is exposed to the DPU once through a cross-GVMI memory key (CGMK) alias; later checkpoints submit only compact (region, offset, size) descriptors, which the DPU bounds-checks before pulling data
  • Data path: DPU-side DMA (MMO) pull from host memory, RDMA relay to the peer DPU, then placement into remote host memory, streamed through bounded DPU buffers
  • Framework integration: VELOC and SCR for CPU-resident HPC state; PyTorch Distributed Checkpoint and ByteCheckpoint for GPU training through double-buffered pinned host staging, without changing their APIs

Key results: up to 16× less extra host CPU time per checkpoint, up to 12× lower checkpoint blocking time, a persistent remote copy up to 4.5× sooner, and <1% end-to-end overhead in PyTorch DCP training.

FitCache I/O flow

FitCache

IPDPS 2026 · Multi tier caching for distributed deep learning I/O

Transparent drop in framework that intercepts POSIX reads and serves training data from the fastest tier among memory, NVMe, and PFS via concurrent fetch.

  • LD_PRELOAD client intercepts open read close and forwards requests to a cache server
  • Hierarchical cache across DRAM or PMem and NVMe above Lustre or BeeGFS
  • Concurrent multi tier fetch returns the earliest responder, tolerating tier variability
  • Scaled evaluation on Frontier up to 2048 GPUs and smaller clusters

Key results: up to 40% training time reduction and up to 71.6% per batch I/O latency reduction (188.3ms to 53.4ms).

Visual chirality effect on hash codes

Visual Chirality in Image Hashing

Neural Computing and Applications 2023 · Data augmentation analysis for deep hashing

This project studies visual chirality, where flipping an image can change its semantics and shift the hash code distribution, which can hurt retrieval when flip augmentation is used.

  • Designed a ResNet 50 based classifier to identify images with chiral cues by comparing predictions of original and flipped images
  • Constructed chiral datasets with controlled proportions of chiral images on VOC2007, MS-COCO, and NUS-WIDE
  • Evaluated three representative hashing pipelines: HashNet, DCH, and BYOL plus DCH under multiple hash code lengths
  • Provided visual explanations via activated region comparisons between original and flipped images

Key results: performance varies strongly with the proportion of chiral data, and the best performance consistently appears when chiral images account for about 15% to 25% or 75% to 85%.