background
to build a private data center for ads training and SIL. a few concepts there:
linux remote desktop
Linux GUI is based on x-protocol, which is kind of CPU based, during deploying gpu-based app in docker swarm, I have studied xserver, ssh -X channels give the way for visualization desktop and simple apps(e.g. glxgears) among xserver and xclient. while for really gpu-densive app, e.g. game rendering, machine learning, ssh -X can’t really use gpu resource well, which needs further support from device management mechanism from docker or k8s.
virutal machine/ cloud desktop
VM is based on hypervisor, known as a virtual machine monitor(VMM), a type of virtualization software that supports the creation and management of VMs, hypervisor translate requests between the physical and virtual resources, making virtualization possible.
A hypervisor allows one host computer to support multiple guest VMs by virtually sharing its resources, like memory and processing.
Generally, there are two types of hypervisors. Type 1 hypervisors, called “bare metal,” run directly on the host’s hardware. Type 2 hypervisors, called “hosted,” run as a software layer on an operating system, like other computer programs, the most common e.g. vmware, citrix.
When a hypervisor is installed directly on the hardware of a physical machine, between the hardware and the operating system (OS), it is called a bare metal hypervisor, which separates the OS from the underlying hardware, the software no longer relies on or is limited to specific hardware devices or drivers.
VDI
a type of vm application is virtual desktop inftrastructure(vdi), VDI hosts desktop environments on a centralized server and deploys them to end-users on request. this process also known as server virtualization
VDI is not necessarily requires a physical gpu, without a gpu, we can still run vmware, but the performance is poor. but for many other VM usage, beyond VDI, we still need ways to access GPU devices in VM.
GPU techs in VM
vwmare has an introduction about gpu tech in vm:
software 3D. using software to mimic GPU calculating. all 3d calculating is done by CPU
vsga(vitual shared graphics acceleration), each virtual desktop has a vsga driver, which has a ESXi module, inside which has a GPU driver, which can call the physical gpu, but this shared mode is through Hypervisor
- vdga(virtual dedicated graphics accleration), or pass through mode, the physical GPU is assigned to a special virtual desktop for one user only.
- vgpu, split physical GPU to a few virtual-GPU, each virtual desk can have one vGPU.
VM vs docker
docker is a light-weight virtualization way, which don’t need hypervisor, and the app in docker is access the host physical devices directly, making docker looks like a process, rather than a virtual OS; and scientfic computing apps run in VM is actually using the virtual CPU from hypervisor, which bypass the cpu optimzation used for math calcualting. so usually a physical machine can start hundres or thousands docker, but can only run a few VM.