gRemote: Cloud rendering on GPU resource pool based on API-forwarding

Evolution of GPU virtualization to resource pooling

“…Traditional GPU workload management schemes typically adopt fixed resource allocation strategies, which cannot adaptively schedule and allocate resources according to the changing characteristics of actual workloads, thereby affecting GPU performance and efficiency [12]. This has led to the problem of underutilization of resources, with these GPU resources usually provided to each client at the granularity of the entire physical GPU [13]. However, in many actual situations, only a small part of the physical GPU is sufficient to complete the task.…”

“…API redirection is one of the popular virtualization methods for GPUs, which intercepts the application's API calls to the GPU and forwards them to the underlying physical or virtual GPU to achieve virtualization [13]. Common API redirection technologies include vDGA and vGPU.…”

“…With the maturity of GPGPU design, the function of GPUs is no longer limited to graphics acceleration. Therefore, remote API redirection systems not only forward rendering commands but also focus on GPGPU-related commands, including rCUDA [13].…”

“…Although this technology can improve GPU utilization, it puts a significant burden on the server since all tasks are executed on the server-side, and resource sharing between multiple applications can exacerbate the server-side pressure. In contrast, remote API redirection systems require cooperation between the server and client machines, and are therefore called cloud-edge integration frameworks [13]. This system can solve some of the server-side pressure problems.…”