Proxmox Sharing GPU For Multiple VMs |

In modern virtualization environments, the demand for GPU resources is growing, especially with the rise of GPU-intensive applications like machine learning, video editing, and gaming. Proxmox, a popular open-source virtualization platform, offers ways to share a single physical GPU among multiple virtual machines (VMs). This capability allows for efficient resource utilization and cost savings by avoiding the need for dedicated GPUs for each VM. This sharing can be achieved through various techniques, including GPU passthrough and virtualization technologies like NVIDIA vGPU.

While GPU passthrough dedicates the entire GPU to a single VM, vGPU enables the creation of virtual functions (VFs) that can be assigned to multiple VMs, allowing concurrent access to the GPU’s resources. Configuring GPU sharing in Proxmox involves several steps, including installing necessary drivers, setting up IOMMU for PCIe passthrough, and creating vGPU profiles. This article provides an overview of the methods and considerations for sharing GPUs with multiple VMs in a Proxmox environment.

Understanding GPU Sharing in Virtualization:

GPU sharing in virtualization allows multiple virtual machines (VMs) or containers to utilize a single physical GPU, optimizing resource utilization and reducing costs. This is particularly important for GPU-intensive applications like AI/ML processing, video editing, cloud gaming, and computational simulations. Several techniques facilitate GPU sharing, each with its own characteristics, benefits, and limitations.

GPU Passthrough vs. GPU Sharing:

1: GPU Passthrough:

This method dedicates an entire physical GPU to a single VM. It provides near-native performance since the VM has direct access to the GPU’s resources. However, it does not allow sharing of the GPU among multiple VMs, leading to underutilization if the VM does not fully utilize the GPU’s capabilities.

2: GPU Sharing (vGPU, MIG, Time-Slicing):

These techniques enable multiple VMs or containers to share a single physical GPU.

• Virtual GPU (vGPU): vGPU technology virtualizes the GPU, creating virtual instances that can be assigned to individual VMs. A hypervisor manages the allocation and scheduling of vGPUs, ensuring fair distribution and optimal utilization. NVIDIA offers vGPU solutions primarily from their Tesla, Quadro, and A100 series, while AMD offers similar virtualization with SR-IOV through their Firepro S-Series.

• Multi-Instance GPU (MIG): NVIDIA’s MIG technology partitions a single physical GPU into multiple isolated GPU instances at the hardware level. Each instance operates independently with dedicated compute, memory, and bandwidth resources, providing performance isolation and security.

• GPU Time-Slicing: This technique divides the GPU’s processing time into discrete slices, allocating a portion of the GPU’s compute and memory resources to different tasks or users sequentially. It enables concurrent execution of multiple tasks on a single GPU, maximizing resource utilization. However, it may introduce overhead due to rapid context switching.

Use Cases for Sharing a GPU in a Virtualized Environment:

1: Virtual Desktop Infrastructure (VDI): vGPU is useful where GPU needs to be made available on a virtual machine, such as virtual desktop infrastructure (VDI).

2: Cloud Gaming: Sharing GPUs allows multiple users to access graphically intensive games via the cloud.

3: AI/ML Workloads: vGPU is important to run AI/ML workloads in containerized environments.

4: Remote Workstations: Sharing GPUs enables remote access to high-performance workstations for tasks like video editing and 3D rendering.

5: Computational Science: GPU virtualization is used in computational science, such as hydrodynamics simulations.

Challenges and Limitations of GPU Sharing:

• Hardware Requirements: vGPU and mediated pass-through require specific GPUs that are compatible with virtualization.
• Isolation: Mediated pass-through may offer limited isolation between VMs when accessing GPU resources.
• Performance Overhead: GPU time-sharing can introduce overhead due to rapid context switching between processes.
• Complexity: Setting up and managing GPU sharing technologies can be complex, requiring specialized knowledge and expertise.
• Compatibility: Not all applications and operating systems are fully compatible with GPU virtualization technologies.
• Resource Management: Efficiently managing and allocating GPU resources among multiple VMs requires careful planning and monitoring.

Hardware and Software Requirements for GPU Sharing in Proxmox:

Achieving effective GPU sharing in Proxmox requires careful consideration of both hardware and software components. The specific requirements depend on the chosen method, such as GPU passthrough, vGPU, or other virtualization techniques.

Hardware Requirements:

GPU:

• GPU Passthrough: Any GPU compatible with Proxmox can be used for passthrough, dedicating the entire GPU to a single VM.

• vGPU (NVIDIA): Requires specific NVIDIA GPUs from the Tesla, Quadro, or A100 series5. These GPUs are designed to support virtualization and the creation of virtual GPU instances.

• vGPU (AMD): AMD FirePro S-Series GPUs support virtualization with SR-IOV.

• Integrated GPU (Intel): Newer Intel iGPUs support SR-IOV, which allows the GPU to be split into multiple virtual devices.

CPU and Motherboard:

• IOMMU Support: The CPU and motherboard must support the Input/Output Memory Management Unit (IOMMU) to enable efficient device isolation and virtualization. Intel VT-d or AMD-Vi technology is necessary for PCIe passthrough to function correctly.

RAM: Sufficient RAM is required to support both the Proxmox host and the virtual machines sharing the GPU. The amount of RAM needed depends on the workload and the number of VMs.

Software Requirements

Proxmox VE: A Proxmox Virtual Environment (VE) is required as the host operating system.

Operating System for VMs: The guest operating systems for the VMs must be compatible with the chosen GPU sharing method. Linux VMs should have a reasonably recent kernel. Note that VirGL, a graphics backend, has no Windows support currently as a driver needs to be written.

Drivers:

• NVIDIA vGPU Drivers: Specific NVIDIA vGPU drivers must be installed on both the Proxmox host and the guest VMs to enable vGPU functionality.

Kernel:

• A newer kernel version may be needed to support SR-IOV for iGPUs2. The Opt-In Kernel 5.19+ is recommended.

VFIO: Proxmox supports VFIO (Virtual Function I/O) passthrough, which allows direct access to physical GPUs from VMs.

To share a GPU between the host and Linux VMs, the following is required:

• Proxmox 7.4
• Opt-In Kernel 5.19+
• AMD GPU is recent enough to use the “amdgpu” driver.
• Linux VM containing a reasonably recent kernel (Virgl has been compatible since the 4.x series and most distributions should be using 5.15+).

On the Proxmox host, execute the following commands:

text# apt update && apt upgrade -y
# apt install libgl1 libegl1

Next, select “virgl” as your display under the Hardware tab in the Proxmox web UI. It’s important to note that when choosing a GPU for sharing, the amount of memory available on the card affects the number of virtual machines you can run. Also, when setting up SR-IOV with iGPUs, performance may be abysmal

Methods for GPU Sharing in Proxmox:

Here’s a detailed look at methods for GPU sharing in Proxmox:

GPU Passthrough (One GPU per VM)

What is PCI passthrough? 

• PCI passthrough allows a virtual machine to directly access a physical PCI device, such as a GPU, as if it were directly connected to the VM. This provides near-native performance since the VM has direct control over the hardware. However, it means that the GPU cannot be shared with other VMs.

Steps to configure PCI passthrough in Proxmox:

• Enable IOMMU: Ensure that your system’s IOMMU (Input-Output Memory Management Unit) is enabled in the BIOS.

• VFIO Passthrough: Proxmox supports VFIO (Virtual Function I/O) passthrough, which allows direct access to physical GPUs from VMs. This provides dedicated GPU resources to each VM.

GPU Sharing with SR-IOV:

• Understanding SR-IOV (Single Root I/O Virtualization): SR-IOV is a PCIe standard that allows a single physical PCIe device to appear as multiple separate physical devices. This enables multiple virtual machines to share a single physical GPU.

• Steps to enable SR-IOV in Proxmox: The process involves installing a newer kernel version and modified dkms modules.

• GPU models that support SR-IOV: Newer integrated GPUs (iGPUs) support SR-IOV, allowing the GPU to be split into multiple virtual devices.

NVIDIA vGPU and AMD MxGPU:

What is NVIDIA vGPU and how does it work? 

• NVIDIA vGPU technology virtualizes the GPU, creating virtual instances that can be assigned to individual VMs. A hypervisor manages the allocation and scheduling of vGPUs, ensuring fair distribution and optimal utilization. vGPU profiles define the amount of memory dedicated to each virtual machine and must be split evenly. When running vGPU, you can only select a single vGPU profile to share out with your virtual machines at a single time.

AMD MxGPU – an alternative to NVIDIA vGPU: 

• AMD MxGPU (now called SR-IOV GPU virtualization) is an alternative to NVIDIA vGPU.

Licensing requirements for NVIDIA vGPU: 

• Note that NVIDIA vGPU often requires specific enterprise GPUs and licensing.

Intel GVT-g (For Intel GPUs)

• How Intel GVT-g enables GPU sharing: Intel GVT-g is a technology that enables GPU sharing on Intel integrated GPUs.

• Setting up GVT-g in Proxmox: To share a GPU between the host and Linux VMs, Proxmox 7.4 and Opt-In Kernel 5.19+ are required4. You also need an AMD GPU recent enough to use the “amdgpu” driver and a Linux VM containing a reasonably recent kernel.

Sharing GPU in Proxmox:

Sharing a GPU in Proxmox allows multiple VMs to utilize a single physical GPU, optimizing resource use. Methods include GPU passthrough, SR-IOV, and vGPU. GPU passthrough dedicates the entire GPU to one VM, offering near-native performance. SR-IOV allows a single GPU to appear as multiple devices, shared among VMs. NVIDIA vGPU creates virtual GPU instances assignable to multiple VMs. Configuration involves enabling IOMMU in BIOS, modifying GRUB, and blacklisting GPU drivers on the Proxmox host. VFIO is configured to allow VMs direct GPU access. Performance optimization includes addressing CPU, RAM, and storage bottlenecks. Tools like

text

nvidia-smi

can monitor GPU usage.

Sharing a GPU requires compatible hardware and drivers. For remote access, Parsec can be used with hardware rendering enabled. Each method has its complexity, compatibility, and performance overhead to consider. Selecting the appropriate method depends on the workload requirements and available hardware.

Additional notes:

• vGPU Profiles: vGPU profiles define the amount of memory dedicated to each virtual machine and must be split evenly.
• Integrated GPU Passthrough: It’s possible to pass through an integrated GPU to a virtual machine.
• Explanation: Some guides offer explanations of what the commands and configurations do.
• Motivation: Sharing existing hardware through virtualization can be a solution when buying new hardware is not feasible.
• Troubleshooting: Testing and troubleshooting are important steps in the process.

Performance Optimization and Troubleshooting:

To optimize performance and troubleshoot GPU sharing in Proxmox, consider these points:

• GPU Passthrough Performance: GPU passthrough gives a VM direct access to a GPU, yielding near-native performance. Ensure other components like storage and CPU don’t bottleneck performance.

• Bottlenecks: Identify and address any performance bottlenecks related to CPU, RAM, or storage. Using an NVMe SSD for VM storage can help.

• GPU Memory: The amount of video memory on your graphics card determines the number of VMs you can run simultaneously when sharing a GPU. NVIDIA’s vGPU profiles define the memory dedicated to each VM.

• Monitoring: Monitor GPU usage and performance using tools like NVIDIA System Management Interface (nvidia-smi) to ensure optimal resource allocation and identify potential issues.

• Remote Access: For remote access and high-fidelity gaming, use software like Parsec and enable hardware rendering. Configure Parsec to use the virtual GPU as the primary display for optimal performance.

• VFIO Configuration: Ensure correct vendor IDs are set in text/etc/modprobe.d/vfio.conf.

• VirGL: The VirGL display setting allows the VM to offload some workloads to the host’s graphics card, which can speed up desktop performance.

Best Practices for GPU Sharing in Proxmox:

Here’s a summary of best practices for GPU sharing in Proxmox, incorporating insights from the search results:

• Hardware Selection: Choose GPUs with virtualization support compatible with Proxmox. AMD and NVIDIA GPUs have varying levels of support for GPU virtualization technologies.

• Enable IOMMU: Ensure the IOMMU (Input-Output Memory Management Unit) is enabled in the BIOS for efficient device isolation and virtualization.

• VFIO Passthrough: Use VFIO (Virtual Function I/O) passthrough, which allows direct access to physical GPUs from VMs, providing dedicated GPU resources.

• Resource Allocation: Balance GPU allocation among VMs to avoid performance degradation or underutilization. High-priority VMs can receive more GPU resources, while lower-priority VMs can have reduced allocation.

• VM Isolation: Maintain isolation between VMs to prevent one VM’s workload from interfering with another’s, which is crucial for stability and performance.

• Monitoring: Utilize monitoring tools within VMs, such as NVIDIA’s System Management Interface (Nvidia-smi) or AMD’s GPU monitoring utilities, to monitor GPU performance and utilization.

• Blacklisting Drivers: Prevent the Proxmox host system from utilizing the GPU by blacklisting the appropriate driver.

• GPU Profiles: When running vGPU, you can only select a single vGPU profile to share with your virtual machines at a single time. vGPU profiles define the amount of memory dedicated to each virtual machine and must be split evenly.

• IOMMU Grouping: Ensure that the GPU is in its own IOMMU group. This is important for a successful GPU passthrough.

• Disable Framebuffer: Disable the framebuffer by adding text video=vesafb:off,efifb:off to the GRUB command line.

• Test and Troubleshoot: Testing and troubleshooting are important steps in the process

Conclusion

Sharing a GPU in Proxmox is a powerful way to optimize resource utilization and enhance the performance of virtualized environments. By understanding the different methods available, such as GPU passthrough, SR-IOV, and vGPU, administrators can tailor their configurations to meet specific workload requirements. Proper planning, including hardware selection, BIOS configuration, and driver installation, is essential for a successful implementation. Performance can be further optimized by monitoring GPU usage, allocating resources effectively, and ensuring proper isolation between VMs. While challenges exist, like compatibility issues and complex configurations, the benefits of GPU sharing, such as cost savings and improved resource utilization, make it a valuable technique for modern virtualization deployments. Continuous monitoring and refinement of configurations are crucial to maintaining optimal performance and stability.

FAQs

1. Can I share a single GPU among Windows and Linux VMs?

Yes, by using VFIO passthrough, you can share a GPU between VMs running different operating systems.

2. What happens if one VM monopolizes GPU resources?

Implementing GPU scheduling policies helps prevent resource monopolization, ensuring fair distribution among VMs.

3. Is it possible to allocate fractions of a GPU to different VMs?

Yes, VFIO passthrough allows you to allocate fractions of GPU resources to VMs, offering fine-grained control.

4. How can I monitor GPU performance and utilization in VMs?

Utilize monitoring tools within VMs, such as NVIDIA’s System Management Interface (Nvidia-semi) or AMD’s GPU monitoring utilities.

5. Are there any security concerns with GPU sharing?

Ensuring proper isolation between VMs and using the latest GPU drivers can mitigate security risks associated with GPU sharing.

Last Updated on 11 February 2025 by Ansa Imran