{"id":42117,"date":"2019-03-18T00:00:00","date_gmt":"2019-03-17T21:00:00","guid":{"rendered":"https:\/\/prohoster.info\/blog\/blog_prohoster\/situatsiya-virtualnye-gpu-ne-ustupayut-po-proizvoditelnosti-zheleznym-resheniyam"},"modified":"2020-02-18T13:44:55","modified_gmt":"2020-02-18T10:44:55","slug":"situatsiya-virtualnye-gpu-ne-ustupayut-po-proizvoditelnosti-zheleznym-resheniyam","status":"publish","type":"post","link":"https:\/\/prohoster.info\/en\/blog\/administrirovanie\/situatsiya-virtualnye-gpu-ne-ustupayut-po-proizvoditelnosti-zheleznym-resheniyam","title":{"rendered":"Situation: virtual GPUs are not inferior in performance to hardware solutions.","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

In February, a conference dedicated to high-performance computing (HPC) took place at Stanford. Representatives from VMware reported that when using GPUs, the system based on a modified ESXi hypervisor performs comparably to bare metal solutions.<\/p>\n

We discuss the technologies that made this possible.<\/p>\n

\"Situation:<\/a><\/noindex><\/a><\/noindex>
\n\/ \u0444\u043e\u0442\u043e Victorgrigas<\/a><\/noindex> CC BY-SA<\/a><\/noindex><\/i><\/p>\n

Performance Issues<\/h2>\n

\nAnalysts estimate that approximately 70% of workloads in data centers are virtualized<\/a><\/noindex>. However, the remaining 30% still operate on bare metal without hypervisors. These 30% mostly consist of high-demand applications, such as those related to training neural networks, and utilize graphics processors.<\/p>\n

Experts explain this trend by stating that a hypervisor, as an intermediate layer of abstraction, can impact the overall performance of the system. Research from five years ago indicates<\/a><\/noindex> a 10% drop in performance. Therefore, companies and data center operators are hesitant to migrate HPC workloads to a virtual environment.<\/p>\n

However, virtualization technologies are evolving and improving. At a conference a month ago, VMware announced that the ESXi hypervisor does not negatively affect GPU performance. The computation speed may decrease by three percent, which is comparable to bare metal metrics.<\/p>\n

How it works<\/h2>\n

\nTo enhance the performance of HPC systems with graphics processors, VMware implemented several changes to the hypervisor. In particular, it was stripped of the vMotion feature. This feature is necessary for load balancing and typically moves virtual machines (VMs) between servers or GPUs. Disabling vMotion means that each VM is now tied to a specific graphics processor. This has helped reduce overhead in data exchange.<\/p>\n

Another key component of the system is the technology<\/a><\/noindex> DirectPath I\/O. It allows the CUDA driver for parallel computing to interact with virtual machines directly, bypassing the hypervisor. When multiple VMs need to run on a single GPU, the GRID vGPU solution is utilized. It segments the memory of the card into several sections (but the computation cycles are not divided).<\/p>\n

The operation scheme of two virtual machines in this case will look as follows:<\/p>\n

\"Situation:<\/p>\n

Results and Predictions<\/h2>\n

\nCompany conducted tests<\/a><\/noindex> hypervisor, training a language model based on TensorFlow<\/a><\/noindex>. The performance \"penalty\" was only 3-4% compared to bare metal. In return, the system gained the ability to allocate resources on demand based on current loads.<\/p>\n

The IT giant also conducted tests<\/a><\/noindex> with containers. Company engineers trained neural networks to recognize images. Resources from one GPU were shared among four container VMs. As a result, the performance of individual machines decreased by 17% (compared to a single VM with full access to GPU resources). However, the number of images processed per second tripled.<\/a><\/noindex> Such systems are expected to find<\/a><\/noindex> application in data analysis and computer modeling.<\/p>\n

Among the potential problems VMware may face, experts highlight<\/a><\/noindex> have a rather narrow target audience. Currently, a small number of companies work with high-performance systems. Although according to Statista, note<\/a><\/noindex>by 2021, 94% of workloads in the world\u2019s data centers will already be virtualized. According to forecasts<\/a><\/noindex> by analysts, the HPC market is expected to grow from $32 billion to $45 billion from 2017 to 2022.<\/p>\n

\"Situation:
\n\/ \u0444\u043e\u0442\u043e Global Access Point<\/a><\/noindex> PD<\/i><\/p>\n

Similar solutions<\/h2>\n

\nThere are several analogs on the market, developed by major IT companies: AMD and Intel.<\/p>\n

The first company for GPU virtualization offers<\/noindex> uses an SR-IOV (single-root input\/output virtualization) approach. This technology provides VMs with access to part of the system's hardware capabilities. The solution allows sharing a GPU between 16 users while maintaining equal performance for the virtualized systems.<\/p>\n

As for the second IT giant, their technology is based<\/a><\/noindex> on the Citrix XenServer 7 hypervisor. It combines the operation of the standard GPU driver and the virtual machine, allowing the latter to display 3D applications and desktops on devices for hundreds of users.<\/p>\n

The future of the technology<\/h2>\n

\nVirtual GPU developers are betting<\/a><\/noindex> on the implementation of AI systems and the growing popularity of high-performance solutions in the business technology market. They hope that the need for processing large volumes of data will increase the demand for vGPU.<\/p>\n

Currently, manufacturers are looking for a way<\/a><\/noindex> combine CPU and GPU functionality in a single core to accelerate graphics-related tasks, perform mathematical computations, logical operations, and data processing. The emergence of such cores in the market will change the approach to resource virtualization and their allocation among workloads in virtual and cloud environments.<\/p>\n

\nWhat to read on the topic in our corporate blog:<\/p>\n