Improving GPU-Accelerated Virtual Machines Startup Performance in Cloud Environments

Cloud-based GPU provisioning is increasingly crucial for AI workloads.However, largememory GPU VMs often experience lengthy boot times due to the memory reservation requirement.Two principal methodologies are commonly employed to mitigate this issue: huge page memory in host operating system and thread context by the underlying hypervisor.This research investigates the impact of huge pages and thread context on mitigating this startup time bottleneck.Our results quantify the performance gains achieved through these optimizations and demonstrate that configuration of thread context and huge pages can significantly reduce VM startup time by 24% and 80.9%, respectively.This work provides practical guidance for improving the deployment of GPU VMs in cloud environments, contributing to faster application startup.