In scenarios where heterogeneous computing services are used ,“ There's an excess of computing resources ” The problem is very common . Take cloud games for example , Businesses usually need only one physical GPU With a fraction of the computing power, you can do graphic or visual computing smoothly . For this kind of application scenarios which require less computing power , Lightweight heterogeneous computing products are more applicable .

In order to enable users to start business with more granular computing resources , Jingdong Zhilian cloud recently launched a new product based on NVDIA vGPU Virtualization of Technology GPU example , With the help of software, the physical... Can be realized through the technology of piecewise virtualization GPU The cards are redivided , After the split vGPU It has the computing power and display memory of the corresponding partition proportion , The same piece of GPU After virtualization, the card can be allocated to different virtual machines . In the cloud, users can select the instance specifications that match the computing power according to the load , Meet a variety of heterogeneous computing scenarios , Lower the cloud GPU Use cost .

Jingdong Zhilian cloud launched vGPU The virtual machine instance contains C type （Virtual Compute Server） and Q type （Quadro vDWS） Two paragraphs , all ** carrying NVIDIA Tesla P40, Support 1/2、1/4、1/6 Three kinds of partition granularity **, And offer a variety of CPU、 Memory configuration combination , Users can choose the right amount of computing resources on demand , Improve system flexibility , And substantially reduce costs .

C type vGPU Examples are mainly for AI、 Computer learning and scientific computing , It is more suitable for design institutes in Colleges and universities 、 Research institutions and other deep learning teaching and experimental scenarios ;Q type vGPU Examples are mainly for real-time rendering of the film and television industry 、 Graphics and image processing and architectural industrial design and other professional image processing scene , Can support Maya、3DMAX、UG、BIM Professional graphics processing software , Satisfy the user to GPU The need for graphic design .

▲C type （Virtual Compute Server）vGPU Example specifications ▲

▲Q type （Quadro vDWS）vGPU Example specifications ▲

stay vGPU Before technology came out , On the cloud GPU Most cloud hosts adopt the direct mode （GPU passthrough）, In through mode GPU Bypass the operating system , As a physical device, it is directly provided to the virtual machine , Because there is no device simulation and conversion process, the performance loss is minimal , It can satisfy most large-scale parallel computing scenarios .

however , Through mode is limited to GPU Physical restrictions on card use , A single virtual machine can carry at least one piece of GPU, Physical servers GPU The number of virtual machines determines the number of virtual machines CPU And memory allocation ratio , For example, most of the running time of a business is to GPU Less computing power is needed , There will be a lot of wasted computing resources . therefore ,GPU Lightweight with low average core usage GPU application , It's a great choice vGPU Specification host for deployment .

• Lightweight model reasoning service

In the context of deep learning , What is needed for online reasoning GPU Resources tend to be less than offline training , But the workload will fluctuate due to the impact of online business , There's a lot of concurrency during peak times . This kind of business deployment can choose the appropriate vGPU Specification host as the smallest deployment unit of cluster , In order to make the cluster computing capacity more in line with the actual demand curve of computing power , Improve GPU Utilization of resources , Optimized cost .

• teaching 、 Development scenarios

Universities and teaching institutions are developing AI Related courses , We need to provide a ride GPU As the basic teaching practice environment , The research direction and professional level of the participants are different , Yes GPU The demand for resource computing power is also different , Apply for different specifications on cloud according to teaching tasks vGPU Virtual machine and GPU Virtual machine , It can meet the resource requirements in various scenarios , It can save teaching resources .

And physics GPU The difference between cards is ,NVIDIA vGPU For different scenarios , There are four types of products available . Each type vGPU The runtime requires the corresponding software authorization （License）, There are also different requirements for the operating system .

besides , Different types of vGPU Products also differ in many functional features , You can go to **NVIDIA Official website （ Please stamp the link ** http://3.cn/15-k06ay）.

In terms of Authorization ,vGPU After the virtual machine runs, it will be configured in advance License Server The server initiates an authorization request , Succeed in getting License Then it will run at standard performance , If you get License Failure will run in limited performance mode until authorized .vGPU Virtual machines consume only when they are running License, When the host is stopped or released ,License Will be License Server Automatic recovery .

Recommended reading ：

• Send you 4 A pithy formula Cloud storage selection is no longer difficult

• Jingdong Zhilian cloud new generation distributed database TIDB Architecture unveils

• Than MySQL fast 839 times ！ Uncover analytical databases JCHDB The veil of mystery

Welcome to click 【 Jingdong Zhilian cloud 】, Learn about the developer community

 More wonderful technology practice and exclusive dry goods analysis

 Welcome to your attention 【 Jingdong Zhilian cloud Developer 】 official account