Virtual GPUs and Modern KVM Virtualization

Architectures, Implementations and Deployment Patterns for Near-Native GPU Access in Virtual Machines

Introduction

1. What You Will Learn
2. How to Read This Book
3. Conventions and Source References

Chapter 1: The GPU Virtualization Problem

1. Why GPUs Are Hard to Virtualize
2. Goals of GPU Virtualization
3. The Evolution Timeline
4. Taxonomy of GPU Virtualization Approaches

Chapter 2: KVM and QEMU: The Modern Linux Hypervisor Stack

1. The Two-Part Architecture
2. KVM Architecture
3. QEMU Architecture
4. The KVM IOCTL Interface in Detail
5. Guest Execution Lifecycle

Chapter 3: IOMMU, VFIO and Device Assignment Foundations

1. IOMMU Fundamentals
2. VFIO Framework
3. PCIe Topology and Isolation
4. Security Implications of IOMMU and VFIO

Chapter 4: PCIe Passthrough: The Original Full-GPU Approach

1. The Passthrough Mechanism
2. Architecture Diagram
3. Configuration and Setup
4. Guest Driver Interaction
5. Performance Characteristics
6. The Passthrough Limitations
7. Production Considerations

Chapter 5: Mediated Devices (mdev): Sharing GPU Time Slices

1. The Mediated Device Framework
2. NVIDIA’s vGPU via mdev
3. Intel GVT-g
4. Mediated Device Management
5. Limitations and Constraints

Chapter 6: virtio: The Foundation of Paravirtualized I/O

1. virtio Design Philosophy
2. The Virtqueue Architecture
3. Device Discovery
4. Transport Options
5. virtio in Linux
6. virtio Performance Characteristics
7. Summary

Chapter 7: VirtIO-GPU: Paravirtualized Graphics in the virtio World

1. VirtIO-GPU Device Model
2. Resource Types and Lifecycle
3. Command Submission
4. virtio-gpu with DRM/KMS
5. virglrenderer Backend
6. Venus Protocol for Vulkan
7. DRM Native Context
8. Performance Characteristics
9. Limitations

Chapter 8: NVIDIA GPU Architecture: What the Hypervisor Must Deal With

1. NVIDIA GPU Hardware Generations
2. GPU Driver Architecture
3. GPU Firmware and GSP
4. PCIe BAR Layout
5. Compute APIs and Context Management
6. GPU Memory Architecture
7. What This Means for Virtualization

Chapter 9: The Architecture of virtio-nvgpu: Bridging virtio and NVIDIA

1. Motivation and Goals
2. Design Philosophy: Forward the Driver ABI, Not the Graphics API
3. High-Level Architecture
4. The virtio Device Specification
5. Design Decisions and Rationale
6. Standardized vs. Vendor-Specific Components
7. Relationship to Other Approaches

Chapter 10: virtio-nvgpu Host Components: Kernel Backend and Device Model

1. Host Kernel Module
2. virtio Backend Device
3. Physical GPU Binding
4. Command Translation Path
5. Memory Management on Host
6. NVIDIA Driver Interaction
7. Error Handling and GPU State
8. QEMU Integration

Chapter 11: virtio-nvgpu Guest Components: Frontend Driver and Device Discovery

1. Guest virtio Frontend
2. Guest NVIDIA Driver Integration
3. Device Lifecycle in Guest
4. Guest Kernel Configuration
5. ABI Profile Loading
6. Limitations of the Guest Driver

Chapter 12: Data Paths and Command Submission: End-to-End Workload Flow

1. Application Launch
2. Guest NVIDIA Driver Path
3. Virtio Queue Submission
4. Host Backend Processing
5. Physical GPU Execution
6. Memory Operations
7. Result Return Path
8. Synchronization
9. Complete Flow Diagram

Chapter 13: Memory Management: The Core Challenge of GPU Virtualization

1. GPU Memory Models
2. Guest-Physical to Host-Physical Translation
3. DMA and Coherency
4. Huge Pages and GPU Memory
5. Memory Isolation and Protection
6. Memory Performance
7. virtio-nvgpu Memory Management Summary

Chapter 14: Performance Engineering: Measuring and Maximizing GPU Throughput

1. Performance Baselines
2. Sources of Overhead
3. CPU Overhead and vCPU Pinning
4. PCIe Bus Utilization
5. Memory Bandwidth
6. Benchmarking Methodology
7. Tuning Strategies
8. Comparison with Bare-Metal GPU Execution

Chapter 15: Security, Isolation and Attack Surfaces

1. Isolation Models
2. DMA Protection
3. Attack Vectors
4. GPU Reset and Fault Handling
5. Multi-Tenant Security
6. Trust Boundaries
7. Supply Chain Considerations
8. Recommendations

Chapter 16: Comparison of GPU Virtualization Approaches

1. Comparison Matrix
2. PCIe Passthrough (VFIO)
3. NVIDIA vGPU
4. NVIDIA MIG
5. virtio-nvgpu
6. virtio-gpu with virglrenderer
7. Intel GVT-g (Deprecated)
8. Intel SR-IOV
9. AMD MxGPU and SR-IOV
10. Summary: Making the Right Choice

Chapter 17: Deployment Architectures: From Dev Workstations to Cloud Data Centers

1. Single-GPU Development Machine
2. Multi-GPU Server with Dedicated VMs
3. Multi-Tenant GPU Cluster
4. AI/ML Workload Patterns
5. Graphics Workloads
6. CI/CD Environments
7. Hybrid Cloud and Edge
8. Container vs. VM Trade-offs for GPU Workloads
9. Production Operational Considerations

Chapter 18: Observability, Tracing and Debugging

1. System-Level Diagnostics
2. virtio Device Diagnostics
3. KVM/QEMU Diagnostics
4. Kernel Tracing with ftrace
5. eBPF for Observability
6. GPU-Specific Tools
7. Network and PCIe Diagnostics
8. Performance Profiling

Chapter 19: Troubleshooting GPU Virtualization Issues

1. Issue: GPU Not Visible in VM
2. Issue: IOMMU Group Contains Unwanted Devices
3. Issue: GPU Reset Fails or Hangs
4. Issue: High CPU Usage or Performance Degradation
5. Issue: NVIDIA Driver Installation Fails Inside Guest
6. Issue: VM Crashes on GPU Access
7. Issue: Multi-Tenant GPU Contentions
8. Issue: Security and Isolation Violations
9. Issue: Host Crash Due to GPU Virtualization
10. General Troubleshooting Methodology

Chapter 20: The Future of GPU Virtualization

1. virtio-nvgpu Maturation Path
2. SR-IOV for NVIDIA GPUs
3. Hardware Assisted Virtualization
4. Container Orchestration and GPU Virtualization
5. Open-Source and Vendor-Neutral Standards
6. AI/ML Workload Implications
7. Security Trends
8. Performance Improvements
9. Cloud and Hyperscaler Adoption
10. Remaining Open Problems
11. Conclusion of Future Outlook

Conclusion

References