The problem of “computing resource surplus” is common in heterogeneous computing services. Take cloud gaming as an example, enterprises typically need only a fraction of the computing power of a physical GPU to complete graphics or visual computing smoothly. Lightweight heterogeneous computing products are more suitable for such application scenarios that require less computing power.

In order to enable users to start businesses with more fine-grained computing resources, JD Zhilian cloud recently launched a virtual GPU instance based on NVDIA vGPU technology. The physical GPU card is redivided by software through slice virtualization technology, and the vGPU after partition has the corresponding proportion of computing capacity and video memory. A GPU card can be virtualized and allocated to different cloud hosts. On the cloud, users can select instance specifications that match the computing force based on the load to meet various heterogeneous computing scenarios and reduce the GPU usage cost on the cloud.

VGPU cloud host instances launched by JD Zhaopin cloud include TWO types of Virtual Compute Server (C) and Quadro vDWS (Q), both of which are ** equipped with NVIDIA® Tesla® P40, supporting three granularity of 1/2, 1/4 and 1/6 **. In addition, the system provides a variety of CPU and memory configurations. Users can select appropriate computing resources as required, improving system flexibility and greatly reducing costs.

C-type vGPU instances are mainly for AI, computer learning, scientific computing and other scenarios, and are mostly suitable for teaching and experimental scenarios of deep learning in design institutes and research institutions of colleges and universities. Q-type vGPU instances are mainly for real-time rendering, graphics and image processing, architectural industrial design and other professional image processing scenes in the film and television industry, and can support Maya, 3DMAX, UG, BIM and other professional graphics processing software to meet users’ requirements for GPU graphics design.

▲ Virtual Compute Server (C) vGPU instance specifications ▲

▲Q Type (Quadro vDWS) vGPU instance specifications ▲

Before the introduction of vGPU technology, most GPU cloud hosts on the cloud use GPU passthrough mode. In passthrough mode, the GPU bypasses the operating system and directly provides VMS as physical devices. As there is no device simulation and conversion process, the performance loss is minimal and it can meet most large-scale parallel computing scenarios.

However, in pass-through mode, each cloud host is equipped with at least one GPU due to the physical usage of the GPU card. In addition, the number of Gpus on a physical server determines the CPU and memory allocation ratio of the cloud host. For example, services require less GPU computing power during most running hours, resulting in a large amount of computing resource waste. Therefore, lightweight GPU applications with low average GPU core usage are suitable for deployment on vGPU hosts.

  • Lightweight model reasoning service

In deep learning scenarios, GPU resources required by online reasoning are usually less than those required by offline training, but the workload fluctuates to a certain extent due to the influence of online services, with a large number of concurrent tasks in peak periods. During service deployment, you can select an appropriate vGPU host as the minimum deployment unit of the cluster based on the workload to better match the actual computing force demand curve, improve GPU resource usage, and optimize the cost.

  • Teaching and developing scenarios

AI related courses in colleges and universities and teaching institutions in, the need for the server as the foundation of carrying the GPU teaching practice environment, staff involved in the curriculum research direction and different business level, the demand for the GPU to calculate force also is not the same, on the cloud according to the teaching task according to the need to apply for different specifications of the vGPU cloud hosting and GPU cloud hosting, It can not only meet the resource needs of various scenarios, but also save teaching resources.

Unlike physical GPU cards, NVIDIA VGpus offer four types of products for different scenarios. Each type of vGPU requires a software License and has different operating system (OS) requirements.

In addition, different types of vGPU products are also different in many features, details can be found on **NVIDIA official website (please click the link ** http://3.cn/15-k06ay).

In terms of authorization mode, the vGPU cloud host sends authorization requests to the preconfigured License Server. After obtaining the License successfully, the vGPU cloud host uses standard performance. If the License fails to be obtained, the vGPU cloud host runs in performance limited mode until authorization is obtained. A vGPU cloud host consumes a License only when the host is running. When the host is stopped or released, the License is automatically reclaimed by the License Server.

Recommended reading:

  • Cloud storage selection is no longer difficult

  • Jingdong Zhilian cloud new generation of distributed database TIDB architecture revealed

  • 839 times faster than MySQL! Uncover the mystery of analytical database JCHDB

Welcome to [JD Zhilian Cloud] to learn about the developer community

More wonderful technical practice and exclusive dry goods analysis

Welcome to [JD Zhilian cloud developer] public account