Work

Breaking the Tradeoff: Elastic and Isolated GPU Sharing with Ghost

Year: 2026

Type: software

Abstract: GVM is a GPU sharing and scheduling research prototype for colocating a latency-critical (LC) vLLM service with a best-effort (BE) diffusion or LLaMA-Factory workload. This repository includes the modified NVIDIA kernel module, CUDA interception layer, applications, baselines, experiment harness, and plotting scripts.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.21926634

Authors: Liu, Yicheng, Qiao, Yifan, Tian, Xia, Xu, Yi, Hong, Tony, Zhou, Yutong, Sun, David, Yang, Shuo, Zhao, Yilong, Shu, Junyi, Xu, Harry, Stoica, Ion, Gonzalez, Joseph, Kumar, Sam

Language: English

Topic: Parallel Computing and Optimization Techniques

Subfield: Hardware and Architecture

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7203485301

Cites:

Cited by:

Related to: 10