V-GPU FOR AI COMPUTING AND PARALLEL COMPUTING

Abstract: Modern Artificial Intelligence (AI) and Machine Learning (ML) workloads are heavily constrained by the high cost and underutilization of physical hardware acceleration. Traditional hardware-level GPU virtualization methods introduce severe hypervisor overhead, require proprietary licensing, and lack automated multi-dataset parallel orchestration. This paper introduces v-gpu, an open-source, containerized virtual GPU orchestration platform designed to automate machine learning pipeline execution using isolated Docker environments. The framework utilizes a custom engine to build deterministic runtime environments (vgpu-worker), auto-provision containerized workspaces, and route data using custom smart ingestion paths. For parallel computing clusters, the system dynamically balances multi-dataset workloads across isolated, load-balanced container groups to maximize computational throughput. Upon job execution, the core engine extracts analytical performance metrics—including Mean Squared Error (MSE), classification accuracy, and F1-score—delivering real-time telemetry to a centralized control dashboard. Experimental evaluations demonstrate that the architecture achieves zero-overhead process isolation, reliable data routing, and rapid infrastructure teardown under peak workloads.

Keywords: Virtual GPU, Containerization, Parallel Computing, Machine Learning Analytics, Infrastructure Automation, Cloud Orchestration.