GPUWeaver
Research on shared GPU interconnects and communication scheduling, published at APSys 2026.
Highlights
- Characterizing contention across GPU fabrics.
- Runtime communication scheduling for LLM applications.
Research
LLM systems move data across PCIe, NVLink, and RDMA fabrics. Concurrent transfers can contend for shared resources, affecting compute progress and host–device data movement.
GPUWeaver studies this contention and introduces a lightweight runtime framework that detects it and regulates communication injection to match application needs.
Publication
Unlocking Software-defined GPU Fabric Scheduling in the LLM Era — Danyang Chen, Yufeng Gu, Yibo Huang, Chengxuan Pei, Peichun Hua, Yang Zhou, and Yunming Xiao. APSys 2026.