It is believed that virtual servers with vGPU are expensive. In a short review, I will try to refute this thesis.
An online search immediately reveals rentals of NVIDIA Tesla V100 supercomputers or simpler servers with powerful dedicated GPUs. Similar services are available, for example, , or . Their monthly cost is measured in tens of thousands of rubles, and I wanted to find cheaper options for OpenCL and / or CUDA applications. There are not so many budget VPS with video adapters on the Russian market, in a short article I will compare their computing capabilities using synthetic tests.
Participants
The list of candidates for participation in the review included virtual servers of hosters , , , ΠΈ . There were no particular problems with obtaining access, since almost all providers have a free trial period. UltraVDS officially does not have a free test, but it turned out to be easy to agree: after learning about the publication, the support staff transferred the amount I needed to order a VPS to my bonus account. At this stage, VDS4YOU virtual machines left the race, because for free testing, the hoster requires you to provide a scan of your ID. I understand that you need to protect yourself from abuse, but for verification, passport details or, for example, linking an account on a social network are quite enough - it is required by 1Gb.ru.
Configurations and prices
For testing, mid-range machines were taken, costing less than 10 thousand rubles per month: 2 cores, 4 GB of RAM, 20-50 GB SSD, vGPU with 256 MB VRAM and Windows Server 2016. Before evaluating the performance of VDS, let's look at their graphic subsystems with an armed eye. Created by the company utility allows you to get detailed information about the hardware and software solutions used by hosters. With its help, you can see, for example, the version of the video driver, the amount of available video memory, as well as data on support for OpenCL and CUDA.
1Gb.ru
GPUcloud
RuVDS
UltraVDS
virtualisation
Hyper-V
OpenStack
Hyper-V
Hyper-V
Computing cores
2*2,6GHz
2*2,8GHz
2*3,4GHz
2*2,2GHz
RAM, GB
4
4
4
4
Drive, GB
30 (SSD)
50 (SSD)
20 (SSD)
30 (SSD)
vGPU
RemoteFX
NVIDIA GRID
RemoteFX
RemoteFX
Video adapter
NVIDIA GeForce GTX 1080 Ti
NVIDIA Tesla T4
NVIDIA Quadro P4000
AMD FirePro W4300
vRAM, MB
256
4063
256
256
OpenCL support
+
+
+
+
CUDA support
β
+
β
β
Price per month (when paid for a year), rub.
3494 (3015)
7923,60
1904 (1333)
1930 (1351)
Payment for resources, rub
no
CPU = 0,42 rub/hour,
RAM = 0,24 rubles / hour,
SSD \u0,0087d XNUMX rubles / hour,
OS Windows = 1,62 rubles / hour,
IPv4 = 0,15 RUB/hour,
vGPU (T4/4Gb) = 7 rubles/hour.
from 623,28 + 30 per installation
no
Test period
10 days
7 days or more by agreement
3 days with monthly billing
no
Of the providers reviewed, only GPUcloud uses OpenStack virtualization and NVIDIA GRID technology. Due to the large amount of video memory (4, 8 and 16 GB profiles are available), the service is more expensive, but OpenCL and CUDA applications will work for the client. The rest of the contenders offer vGPUs and smaller VRAMs built using Microsoft RemoteFX. They are much cheaper, but only support OpenCL.
Performance testing
GeekBench 5
With this popular you can measure the performance of the graphics subsystem for OpenCL and CUDA applications. The chart below shows a summary result, while more detailed data for virtual servers , GPUcloud ( ΠΈ ), ΠΈ available on the website of the benchmark developer. If you open them, you will find an interesting fact: GeekBench shows VRAM volumes that are much higher than the ordered 256 MB. The clock frequency of the central processors may also be higher than stated. In virtual environments, this is a common occurrence - a lot depends on the workload of the physical host on which the VPS is running.
Shared "server" vGPUs are weaker than performance "desktop" video adapters when used for heavy graphics applications. Such solutions are intended mainly for computational problems. Other synthetic tests have been carried out to evaluate their effectiveness.
FAHBench 2.3.1
For a comprehensive analysis of vGPU computing capabilities is not suitable, but it can be used to compare the performance of video adapters of different VPS in complex calculations using OpenCL. Distributed Computing Project solves the narrow problem of computer simulation of the folding of protein molecules. Researchers are trying to understand the causes of protein-related pathologies: Alzheimer's and Parkinson's diseases, mad cow disease, multiple sclerosis, and so on. Measured with the utility they created the performance of single and double precision calculations is shown in the diagram. Unfortunately, the utility generated an error on the UltraVDS virtual machine.
Next, I will compare the results of calculations for the dhfr-implicit modeling method.
SiSoftware Sandra 20/20
Plastic bag great for evaluating the computing capabilities of virtual video adapters of various hosts. The utility contains General Purpose Computing Benchmarks (GPGPU) and supports OpenCL, DirectCompute, and CUDA. To begin with, a general assessment of different vGPUs was made. The chart shows a summary result, more detailed data for virtual servers , GPUcloud () and available on the website of the benchmark developer.
There were also problems with the "long" Sandra test. For VPS provider GPUcloud, it was not possible to conduct an overall assessment using OpenCL. When choosing the appropriate option, the utility still worked through CUDA. This test also failed for the UltraVDS machine: the benchmark stopped at 86%, trying to determine the memory latency.
In the general test suite, you cannot see indicators with a sufficient degree of detail or perform calculations with high accuracy. I had to run several separate tests, starting with determining the peak performance of the video adapter using a set of simple mathematical calculations using OpenCL and (if possible) CUDA. It also shows only the overall score, and the detailed results for VPS from , GPUcloud ( ΠΈ ), ΠΈ available on the site.
To compare the speed of encoding and decoding data, Sandra has a set of cryptographic tests. Detailed results are available on the website for , GPUcloud ( ΠΈ ), ΠΈ .
Parallel financial calculations require a supportive double-precision calculation of the adapter. This is another important area of ββapplication for vGPUs. Detailed results are available on the website for , GPUcloud ( ΠΈ ), ΠΈ .
Sandra 20/20 allows you to test the possibilities of using vGPU for scientific calculations with high accuracy: matrix multiplication, fast Fourier transform, etc. Detailed results are available on the website for , GPUcloud ( ΠΈ ), ΠΈ .
Finally, the vGPU image processing capabilities were tested. Detailed results are available on the website for , GPUcloud ( ΠΈ ), ΠΈ .
Conclusions
The GPUcloud virtual server performed well in the GeekBench 5 and FAHBench tests, but Sandra did not rise above the overall level in the benchmark tests. It costs much more than competitors' services, but it has a much larger amount of video memory and supports CUDA. In tests Sandra with high accuracy of calculations, VPS from 1Gb.ru was in the lead, but it is also not cheap and in other tests it proved to be average. UltraVDS turned out to be a clear outsider: I donβt know if there is a connection here, but only this hoster offers AMD video cards to customers. In terms of price / performance ratio, the RuVDS server seemed to me the best. It costs less than 2000 rubles a month, while the tests passed quite well. The final standings look like this:
Place
Hoster
OpenCL support
CUDA support
High performance by GeekBench 5
High performance by FAHBench
High performance by Sandra 20/20
Moderate price
I
RuVDS
+
β
+
+
+
+
II
1Gb.ru
+
β
+
+
+
+
III
GPUcloud
+
+
+
+
+
β
IV
UltraVDS
+
β
β
β
β
+
I had some doubts about the winner, but the review is focused on budget VPS with vGPU, and the RuVDS virtual machine costs almost half the price of the nearest competitor and more than four times the most expensive offer from the reviewed. It was also difficult to share the second and third places, but here, too, the price outweighed other factors.
As a result of testing, it turned out that entry-level vGPUs are not so expensive and you can already use them to solve computing problems. Of course, using synthetic tests it is difficult to predict how a machine will behave under a real load, besides, the ability to allocate resources directly depends on the neighbors in the physical host - make allowances for this. If you find other budget VPS with vGPU on the Runet, feel free to write about them in the comments.
Source: habr.com