-
Notifications
You must be signed in to change notification settings - Fork 95
Old Materials
2014 USENIX conference paper - By Tian, Kevin; Dong, Eddie; David Cowperthwaite.
**Conference link: **https://www.usenix.org/conference/atc14/technical-sessions/presentation/tian
Attachment: A Full GPU Virtualization Solution with Mediated Pass-Through
2014 KVM Summit talk by Song, Jike.
**Conference Link: **
http://www.linux-kvm.org/page/KVM_Forum_2014
http://www.linux-kvm.org/wiki/images/f/f3/01x08b-KVMGT-a.pdf
Attachment: KVMGT: a Full GPU Virtualization Solution
Media processing is increasingly important, which accounts for 60%+ internet traffic. This session will introduce the media cloud solution, based on the latest Intel GPU Virtualization Technology and OpenStack cloud software. It will be a big IA differentiator, in the industry move to cloud based services.
Attachment: Media Cloud Based on Intel® GVT-g and OpenStack - Chinese Version Attachment: Media Cloud Based on Intel® GVT-g and OpenStack - English Version
2014 IDF (Intel Developer Forum) presentation by Tian, Kevin.
Agenda:
•Why GPU Virtualization?
•The Way to Full GPU Virtualization
•Architecture Overview
•Key Techniques
•Summary
Attachment: XenGT: A High Performance Graphics Virtualization Solution on Intel® Processor Graphics - Chinese Version Attachment: XenGT: A High Performance Graphics Virtualization Solution on Intel® Processor Graphics - English Version
This foil introduces how GVT-g can support Media Cloud usages, for example video conference, media encoding/transcoding. GVT-g can achieve ~90% performance of native H.264 transcoding.
-
What's GPU virtualization
-
Media processing opportunity
-
Media Cloud examples
-
Video conference usage case
-
Media Cloud building blocks
-
Performance evaluation
Attachment: intel_graphics_virtualization_for_media_cloud.pdf
In 2014 LinuxCon, Intel introduced an idea and implementation about XenGT: A Full GPU Virtualization Solution with Mediated Pass-Through.
Conference link: http://events.linuxfoundation.org/sites/events/files/slides/XenGT-LinuxCollaborationSummit-final_1.pdf
Attachment: 2014_xengt-linuxcollaborationsummit-final_1.pdf
This is a presentation talk in GPU Virtualization Workshop, under Open Source Operating System Annual Technical Conference 2015, Beijing, China.
The presentation starts from a background introduction of GPU virtualization, then focus on Intel's Full GPU Virtualization with Mediated Pass-Through technology and its implementation on Xen and KVM hypervisor. Last, it gives the audience a project status update of XenGT/KVMGT as of Q3'15.
Attachment: GVT-g: A Full GPU Virtualization Solution with Mediated Pass-Through
This is one page introduction about Intel GVT-g technology.
Attachment: gvt-g_introduction_-_broachure.pdf
[2015 USENIX Annual Technical Conference] The increasing adoption of Graphic Process Unit (GPU) to computation-intensive workloads has stimulated a new computing paradigm called GPU cloud (e.g., Amazon’s GPU Cloud), which necessitates the sharing of GPU resources to multiple tenants in a cloud. However, state-of-the- art GPU virtualization techniques such as gVirt still suffer from non-trivial performance overhead for graphics memory-intensive workloads involving frequent page table updates.
To understand such overhead, this paper first presents GMedia, a media benchmark, and uses it to analyze the causes of such overhead. Our analysis shows that frequent updates to guest VM’s page tables causes excessive updates to the shadow page table in the hypervisor, due to the need to guarantee the consistency between guest page table and shadow page table. To this end, this paper proposes gHyvi1, an optimized GPU virtualization scheme based on gVirt, which uses adaptive hybrid page table shadowing that combines strict and relaxed page table schemes. By significantly reducing trap-and-emulation due to page table updates, gHyvi significantly improves gVirt’s performance for memory-intensive GPU workloads. Evaluation using GMedia shows that gHyvi can achieve up to 13x performance improvement compared to gVirt, and up to 85% native performance for multithread media transcoding.
Download URL including presentation video: https://www.usenix.org/conference/atc15/technical-session/presentation/dong
Attachment: boosting_gpu_virtualization_performance_with_hybrid_shadow_page_tables.pdf
From the exponential growth of video on the Internet to desktop virtualization initiatives, media-rich workloads represent a growing share of network traffic. In addition, cloud computing models are incorporating robust media. These usages represent opportunities for businesses to driver new revenue streams and reduce overall costs, but only if the media processing can be managed efficiently. Efficiency in today’s world often implies the use of cloud computing, bug media workloads have traditionally been difficult to schedule in the cloud due to an inability to access graphics processing unit (GPU) offload capabilities for optimal performance. In response to this challenge, graphics virtualization techniques have evolved to allow a media-optimized workload to run on top of a virtualized environment, such as virtualizing the graphics processor for concurrent use to provide direct, dedicated access to a GPU for a single virtualized workload, or providing a pass-through for direct, shared access to the GPU for a number of virtualized workloads.
Attachment: GVT-g flyer for usages
[2016 USENIX Annual Technical Conference] With increasing GPU-intensive workloads deployed on cloud, the cloud service providers are seeking for practical and efficient GPU virtualization solutions. However, the cutting-edge GPU virtualization techniques such as gVirt still suffer from the restriction of scalability, which constrains the number of guest virtual GPU instances.
This paper introduces gScale, a scalable GPU virtualization solution. By taking advantage of the GPU programming model, gScale presents a dynamic sharing mechanism which combines partition and sharing together to break the hardware limitation of global graphics memory space. Particularly, we propose three approaches for gScale: (1) the private shadow graphics translation table, which enables global graphics memory space sharing among virtual GPU instances, (2) ladder mapping and fence memory space pool, which allows the CPU to access host physical memory space (serving the graphics memory) bypassing global graphics memory space, (3) slot sharing, which improves the performance of vGPU under a high density of instances.
The evaluation shows that gScale scales up to 15 guest virtual GPU instances in Linux or 12 guest virtual GPU instances in Windows, which is 5x and 4x scalability, respectively, compared to gVirt. At the same time, gScale incurs a slight runtime overhead on the performance of gVirt when hosting multiple virtual GPU instances.
Download URL: https://www.usenix.org/node/196262
Attachment: gscale_scaling_up_gpu_virtualization_with_dynamic_sharing_of_graphics_memory_space.pdf
[Xen Summit 2016]
In the talk, we gave a demo of our vGPU live migration, then we discussed the GVT-g live migration features, challenges, designs and implementations.
URL: http://events.linuxfoundation.org/events/xen-project-developer-summit
Attachment: xengt-livemigration_1.00.pdf
[Xen Summit 2016]
In this Xen Summit talk, we talked our detailed implementations about how to improve vGPU scalability by using dynamic resource sharing in GVT-g project.
URL: http://events.linuxfoundation.org/events/xen-project-developer-summit
Attachment: gscale_1.00.pdf
The need for dedicated graphics processing, once exclusive to high-end design or manufacturing engineers, is now pervasive. Even common modern applications, such as web browsers and office productivity software, tax the graphics capabilities of the systems on which they run. At the same time, as users have become increasingly mobile, they expect a responsive user experience on any device, anywhere... they go. They also want an environment that allows them to easily and effectively share and collaborate on projects.
All of these needs—mobility, an outstanding user experience, and seamless collaboration—require unencumbered access to data. To provide that access with simpler workloads, IT organizations might implement centralized solutions that focus on security, efficiency, and simplified access. However, some IT pros and users are skeptical about centralized solutions when it comes to graphics. Historically, multiple remote workers using graphics-intensive applications from shared system resources could have a detrimental impact on performance.
Without a solution, employees can resort to collaborating by exchanging files through e-mail, FTP servers, USB drives, or even by traveling to work on projects with colleagues face to face. These practices undermine productivity and create even greater workflow delays.
It doesn’t have to be this way. Graphics virtualization technology, which is available today, can turn skeptics into believers because it lets employees work faster and collaborate in real-time while eliminating the dependency on, and the incremental cost of, expensive dedicated graphics cards.
Reference: www.intel.com/citrix
Attachment: Virtualized Graphics Processing Enhances Collaboration
By running virtual machines (VM) on the Citrix XenServer 7* hypervisor, each app and desktop receives graphics acceleration benefits from Intel’s virtual GPU, Intel® Graphics Virtualization Technology (Intel® GVT-g). Based on Iris Pro graphics, this technology integrates a standard Intel GPU driver within the centralized VM, which enables the VM to render 3D apps and desktops. You can deliver grap...hics-rich applications with cost models that support broad delivery to hundreds or even thousands of users.
Reference: www.intel.com/citrix
Attachment: intel-citrix-vdesktop-refresh.pdf
With Intel GVT-g (KVMGT) benn upstreamed from kernel 4.10, Intel has updated the GVT-g architecture to upstream friendly.
This document highlights the key difference of new architecture, for example, resource mangement, interrupt, GGPTT, scheduling, display. For more details, please check out the latest upstream GVT-g code.
P.S. XenGT upstream is working in progress.
Attachment: Intel GVT-g New Architecture Introduction
Intel's XenGT talk in Xen Summit 2017.
In this talk, we introduced XenGT new architecture and gave 3 options for XenGT upstream, to collect Xen Commuinity's feedback.
Attachment: XenGT from production to upstream, talk in Xen Summit 2017