Shulou(Shulou.com)12/24 Report--

CTOnews.com, November 24, recently, the virtualization technology research and development team of Kirin Software Research and Development Center launched virtual GPU solution V2.0, which uses vhost-user-gpu, 3D rendering synchronization optimization, cloud desktop refresh based on dynamic data and other technologies to improve the 3D rendering performance of virtual machine GPU to more than 80% of the back-end GPU 3D rendering performance.

Kirin Software said that the program does not need professional graphics card support, through consumer graphics cards to enable the virtual machine to meet the needs of image processing, graphic design, 3D modeling and games, while achieving the flexibility and isolation of 1VR N, greatly reducing the user cost.

CTOnews.com attached the technical block diagram of Kirin software virtual GPU solution version V2.0:

According to reports, with the development and wide application of cloud desktop technology, the demand of users for image processing, graphic design, 3D modeling and games on cloud desktop is increasing, which puts forward higher and higher requirements for image rendering ability of cloud desktop.

In May 2022, Kirin Software developed its own virtual GPU solution V1.0 based on open source virtio-gpu paravirtualized graphics card technology and api forwarding scheme. This solution deeply optimizes the virtio-gpu virtual graphics card, and can improve the rendering performance of virtio-gpu to 60% of the back-end GPU performance with back-end GPU support (using glmark2 rendering benchmark).

However, compared with the GPU hardware virtualization scheme, Kirin software virtual GPU solution V1.0 still has a relatively high rendering and forwarding process overhead, and the back-end physical GPU is not fully utilized, resulting in a nearly 40% rendering performance loss. In order to solve this problem, the virtual GPU scheme is further upgraded and optimized, and the Kirin software virtual GPU scheme V2.0 is launched, which reduces the performance loss in the rendering process to less than 20%, and the 3D comprehensive rendering performance can reach more than 80% of the back-end GPU performance.

Kirin Software said that GPU solution V2.0 uses a new technology framework, using vhost-user-gpu virtual graphics card (implemented based on vhost-user technology) to replace the virtio-gpu virtual graphics card used in V1.0 scheme. The vhost-user-gpu virtual graphics card does not need to go through the host kernel kvm module and user mode switching, and can communicate with the virtual machine directly through shared memory, which reduces a lot of user mode (vm exit and vm entry) switching overhead. V1.0 version has less extra overhead and better performance.

In addition, GPU V2.0 also breaks through the rendering command execution real-time monitoring technology and fast command feedback technology to achieve real-time monitoring in the rendering command execution process, improve the command execution monitoring accuracy, and reduce the idle wait of the physical GPU; before the completion of command execution, prepare feedback data in advance to improve feedback speed and reduce feedback time overhead. It realizes the change from complete desktop refresh to dynamic region refresh, that is, only refreshing the areas where the desktop display data has changed, reducing the extra overhead caused by reading the unchanged image data in the physical GPU when the cloud desktop is refreshed.

In order to verify the effectiveness of the scheme, the 3D rendering performance of vhost-user-gpu is tested based on three different graphics cards from Intel set Show, AMD RX 550and Nvidia RTX 2060, as well as two different benchmark programs glmark2 and gputest. The results are as follows:

Under the glmark2 benchmark, when vhost-user-gpu uses three different types of GPU as the back-end rendering acceleration GPU, the test score reaches or exceeds 80% of the original test score of the physical GPU.

The ▲ glmark2 benchmark performance test data compares the scores of vhost-user-gpu and physical GPU in seven different test scenarios under the gputest benchmark. The results show that, in most cases, the performance of the optimized vhost-user-gpu can also reach more than 80% of the original performance of the physical GPU.

▲ GPUTEST benchmark performance data table in addition, Kirin Software Virtualization GPU solution V2.0 also carries out a wide range of architecture adaptation and graphics card adaptation. Under the commercial x86 and arm architecture, the scheme works normally when using the mainstream graphics card manufacturer GPU as the back-end GPU.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.