Π article when we talked about our new service with a video card, we have not touched on some interesting aspects of using virtual servers with video adapters. It's time to complete the testing.
For the use of physical video adapters in virtual environments, we have chosen the RemoteFX vGPU technology, which is supported by the Microsoft hypervisor. At the same time, processors with SLAT support (EPT from Intel or NPT / RVI from AMD) must be installed on the host, as well as video cards that meet the requirements of the creators of Hyper-V. In no case should you compare this solution with desktop adapters in physical machines, which usually show better performance when working with graphics. In our testing, the vGPU will compete with the central processor of the virtual server - quite logical for computing tasks. We also note that in addition to RemoteFX, there are other similar technologies, for example NVIDIA Virtual GPU - it allows you to transfer the graphic commands of each virtual machine directly to the adapter without their translation by the hypervisor.
Tests
The tests used a machine with 4 cores at 3,4 GHz, 16 GB of RAM, a 100 GB solid state drive (SSD) and a virtual video adapter with 512 MB of video memory. The physical server is equipped with NVIDIA Quadro P4000 professional video cards, and the guest system is running Windows Server 2016 Standard (64-bit) with the standard Microsoft Remote FX video driver.
βGeekBench 5
For a start the latest version of the utility , which allows you to measure system performance for OpenCL applications.
We used this benchmark in the previous article and it only confirmed the obvious - our vGPU is weaker than high-performance desktop video cards for solving typical "graphics" tasks.
βGPU Caps Viewer 1.43.0.0
Created by the company the utility cannot be called a benchmark. It does not contain performance tests, but allows you to get information about the hardware and software solutions used. Here you can see that our virtual machine with vGPU only supports OpenCL 1.1 and does not support CUDA, despite the NVIDIA Quadro P4000 video adapter installed in the physical server.
βFAHBench 2.3.1
from the distributed computing project is devoted to solving a highly specialized problem of computer simulation of the folding of protein molecules. This is necessary to study the causes of pathologies associated with defective proteins - Alzheimer's and Parkinson's diseases, mad cow disease, multiple sclerosis, etc. Utility cannot comprehensively evaluate the processing power of a virtual video adapter, but allows you to compare CPU and vGPU performance in complex calculations.
The performance of computations on vGPU using OpenCL, measured using FAHBench, turned out to be about 6 times (for the implicit simulation method - about 10 times) higher than similar indicators for a sufficiently powerful central processor.
Next, we present the results of calculations with double precision.
βSiSoftware Sandra 20/20
Another universal package for diagnosing and testing computers. It allows you to study in detail the hardware and software configuration of the server and contains a huge number of different benchmarks. In addition to CPU computing, Sandra 20/20 supports OpenCL, DirectCompute and CUDA. We are primarily interested in the included in the free version general purpose computing benchmark suites (GPGPU) using hardware accelerators.
pretty good, although they are slightly lower than expected for the NVIDIA Quadro P4000 video adapter. The overhead of virtualization probably affects.
Sandra 20/20 has a similar set of CPU benchmarks. Let's run them to with vGPU computing.
The advantages of the video adapter are clearly visible, however, the settings of the general test package are not completely identical, moreover, the results cannot be seen with the required level of detail. We decided to run several separate tests. At first peak vGPU performance with a set of simple math calculations using OpenCL. essentially similar to the multimedia (not arithmetic!) test Sandra for CPU. For comparison, we place on the same diagram multimedia test of the VPS CPU. Even a CPU with four processing cores noticeably loses to vGPU.
From synthetic tests, let's move on to practical things. Cryptographic tests helped us determine the speed of encoding and decoding data. Here's a comparison of the results for ΠΈ also showed a clear advantage of the accelerator.
Another area of ββapplication for vGPU is financial analysis. Such calculations are easy to parallelize, but they require a video adapter that supports double precision calculations. And again, the results speak for themselves: powerful enough outright loses .
The last test we conducted was scientific calculations with high accuracy. did better again with matrix multiplication, fast Fourier transform and other similar problems.
Conclusions
vGPUs are not well suited for running graphics editors, as well as 3D rendering and video processing applications. Desktop adapters handle graphics much better, but a virtual one can perform parallel computing faster than a CPU. For this, we must say thanks to the productive RAM and more arithmetic logic modules. Collection and processing of data from various sensors, analytical calculations for business applications, scientific and engineering calculations, analysis and billing of traffic, work with trading systems - there are a lot of computing tasks for which GPUs are indispensable. Of course, you can assemble such a server at home or in the office, but you will have to pay a tidy sum for the purchase of hardware and the purchase of licensed software. In addition to capital costs, there are operating costs for maintenance, including electricity bills. There is depreciation - equipment wears out over time, and becomes obsolete even faster. Virtual servers do not have these drawbacks: they can be created as needed and removed when the need for computing power disappears. Paying for resources only when they are needed is always profitable.
Source: habr.com