The Research And Realization Of GPU Qos Technology For Virtual Machine Based On GVirt

Due to the limitation of graphics hardware resources and the high requirements of graphics memory of graphic program for GPU computing resources,it is difficult for compute's GPU to meet the requirements of multiple virtual machines.Especially once the different virtual machines need different GPU computing resources,the modern GPU virtualization technology will not support the on-demand service quality allocation.This thesis got the inspiration from the QoS technology in the field of computer network,adopted the Intel gVirt technology,which is one of the most advanced Intel GPU virtualization technology.However,in gVirt,the GPU time slice scheduling followed round robin style,therefore the GPU computing resources of each virtual machine are equal,it is not good for on-demand service quality allocation.This thesis took the time slice scheduling as the breakthrough point,first confirmed the accuracy of time scheduling through the analysis of the popular GPU benchmark context switch execution time distribution.Then allocated the corresponding proportion of time slice of the realized the QoS feature according to the cap setting of virtual machine in each given period.Finally,divided one period into multiple time slots,allocated the time slice for the first time slot,then adapted and guaranteed the accuracy of QoS feature by reallocating the time slice with accumulation or supplement according to the previous executed time slice from the second time slot.In this way,it implemented the dynamic and flexible GPU computing resources allocation,which achieved the on-demand allocation.This kind of dynamic allocation is automatic that user cannot feel it,only a simple parameter is needed for user to enable the GPU QoS function,it is simple and easy to use.This thesis adopted the solution with using GPU virtualization QoS technology to resolve the GPU computing resources allocation issue,which has validated by Intel GPU virtualization validation team through the QoS weight parameters,QoS performance monitoring,QoS GPU usage monitoring and QoS stability validation and verification based on the most advanced Intel platform and the most popular GPU test tools.The test results show that it allows the GPU computing resources of virtual machines allocated by user freedom and the GPU performance of virtual machines is also more stable than before.At the same time,there is no other negative impact be found in functionality,performance,stability during the validation process.