vMiG: A Fast and NUMA-Aware Virtualized GPU Migration System via Pipelined Zero-Copy Data Paths

GPUs have become indispensable infrastructure for cloud platforms, powering workloads from large-model inference to real-time rendering, and the transition to cloud-native GPU computing is accelerating. One of the key challenges in this transition is live migration, which is essential for enabling transparent workload relocation for load balancing, fault tolerance, and resource consolidation. However, GPU virtualization introduces a critical migration bottleneck rooted in the mediated pass-through (MPT) migration stack: massive device state, particularly VRAM footprints scaling into tens of gigabytes, must be exported through vendor-controlled interfaces whose buffering, copying, and placement behavior limits high-throughput VRAM streaming in practice. To address this challenge, we propose vMiG , a framework that first characterizes why the MPT migration data path is poorly matched to high-throughput VRAM streaming. vMiG then uses this analysis to restructure the migration pipeline for higher VRAM transfer throughput and shorter blackout time, through zero-copy data movement, end-to-end pipelining, and non-uniform memory access (NUMA)-aware control path placement. Evaluations show that NUMA-aware controller placement improves VRAM migration throughput by 58.8%, while copy-path optimizations reduce memory copies from seven to near the theoretical minimum of four. Together, these optimizations achieve VRAM migration throughput up to 5.19 × the baseline rate and reduce end-to-end migration time by up to 58% for VRAM-intensive AI workloads.