怎么合并W2012R2下的4个GRID K1单元的性能

你能解释一下你要做什么，因为这会影响答案。
你打算在服务器上运行什么？
RDSH，Citrix XenApp，还有别的吗？
顺便说一句，将4 K1卡安装到服务器机箱中是不可能的，也没有OEM提供它。
或者你指的是一张卡上的4个GPU？



以下为原文

Can you explain what you're trying to do as that will affect the answer.

What are you planning to run on the server? RDSH, Citrix XenApp, something else?


BTW, it's not possible to fit 4 K1 cards into a server chassis and no OEM offers it. OR are you referring to the 4 GPU's on a single card?

亲爱的杰森，
非常感谢回复，我试着说清楚：RDSH用于基于会话的远程桌面，而不是一台服务器中的4 K1卡。
一台服务器上运行一块GRID K1卡，运行W2012，设备管理器中有4台图形设备 - 与该卡上的四块GPU相关，连接用户（远程桌面）只使用系统中的四块图形卡中的一块。
如果我能以某种方式统一并使用所有4 GPU的性能，我会很高兴，现在只有4中的一个忙。
我没有找到任何方法如何做到这一点，我认为只有解决方案是使用一些Quadro K系列卡，为此应用程序。
问候
MIB



以下为原文

Dear Jason,

thanks a lot for reply, I try to make it clear: RDSH for session based remote desktop, not 4 K1 cards in one server. one GRID  K1 card in one server, W2012 running on it, and in device manager is 4 graphic devices - related to four GPUs on that sigle card, connected users(remote desktop) are using only one of this four graphic cards in system. I would be happy if I could somehow unify and use performance of all 4 GPU, now only 1 of 4 is busy. I didnt find any way how to do it, and I think that only solution is use some Quadro K series card, for this application.

regards

MiB

您所看到的是操作系统处理工作负载放置的方式。
操作系统的一个限制是，它只会将图形工作负载放在它看作主要GPU的位置，因此是第一个K1 GPU。
您看到的GPU是否以100％的速度运行？
您还应该知道，除非您正在运行RemoteFX，否则远程会话将无法看到GPU加速。
Citrix有一些实验性代码允许在GPU上循环使用DirectX工作负载，但这需要XenApp。
如果您可以选择支持直通（XenServer，vSphere或KVM）的虚拟机管理程序，那么您可以配置4个虚拟机，每个虚拟机都有一个GPU。
虽然Hyper-V不支持这一点，但它有一个替代解决方案，但微软不支持它用于RDSH
如果这些不是一个选项，你不能虚拟化，那么你是正确的，你需要一个大小合适的Quadro卡。



以下为原文

What you're seeing is the way that the OS is handling the placement of the workloads. It's a limitation in the OS that it will only place the graphics workload on what it see's as the primary GPU, so the first of the K1 GPU's.

Is the GPU you're seeing load on running at 100%?

You should also be aware that unless you're running RemoteFX, the remote session will not see GPU acceleration.

Citrix have some experimental code for allowing round robin use of DirectX workloads across the GPU's, but that will require XenApp.  

If you have the option of a hypervisor that supports passthrough (XenServer, vSphere or KVM) then you could configure 4 VM's each with one GPU. Hyper-V doesn't support this though, it has an alternate solution but Microsoft doesn't support it for RDSH

If these aren't an option and you can't virtualise then you're correct, you'll need a good sized Quadro card.

我也有兴趣结合两个网格启用卡的功能。
它可能是K或M系列，没关系，因为我目前正在考虑购买一台带有2张卡的服务器，但我怀疑它不能为我的应用和需求提供足够的渲染能力。
基本上，我正在研究大型文件来创建科学可视化并使用Paraview，Catalyst进行现场可视化，我在戴尔T620服务器上安装了Quadro K5200，至少我希望从支持Grid的卡获得相同水平的功率
有远程可视化支持，但根据他们的规格，单卡很难获得足够的性能。
无论如何，你能用你的经验拓宽我。
你有什么建议吗？
问候，
--ufuk



以下为原文

I am also interested in combining the power of two grid enabled card. It could be K or M series, it does not matter because i am currently considering to buy a server with 2 cards on it but i have some doubt that it won't deliver enough rendering power for my application and needs. Basically, i am working on large files to create scientific visualisations and performing in-situ visualisation using Paraview, Catalyst and i had a Quadro K5200 installed on Dell T620 server and at least i am expecting to get same level of power from the Grid enabled cards with remote visualisation support but according to their specs, it is hard to get enough performance from single card. Anyway, could you widen me with your experience. Do you have any suggestion about it.

Regards,

--ufuk

你好
当您查看规格时，M60的性能应与K5200非常相似，尽管相比之下它缺少一些CUDA内核和一些内存带宽，但较新的架构和增加的时钟速度可能会弥补它。
但是，它会在2个会话中执行此操作，而K5200仅支持1个会话。
目前，我不知道任何操作系统可以同时使用M60上的两个GPU来访问完整的16GB，我很高兴被证明是错的，但我确信有一些聪明的人正在努力
这个在哪里... :-)
对于它的价值，我认为这不是一个公平的比较，因为M60是为多个会话而设计的，K5200是单个会话（除非你将它添加到RDS / XenApp服务器......）。
您可能最好不要尝试使用这种型号的GPU，而不是打算使用M5000 / M6000，或者等待更长时间，因为Pascal GPU已经发布并且看了P5000 / P6000（
我相信这些将在10月初开始上市，这是你提到的K5200的继承者。
您可以使用与特斯拉（M60）完全相同的方式将Quadro卡添加到VM，但如上所述，它们是1-1映射，但如果您愿意，仍然可以远程访问它们。
问候
本



以下为原文

Hi

An M60 should be very similar in performance to the K5200 when you look at the specs, although it lacks a few CUDA cores and some memory bandwidth by comparison, but the newer architecture and increased Clock speeds might make up for it. However, it will do that for 2 sessions, whereas the K5200 will only support 1 session. Currently, I'm not aware of any Operating System that can use both GPUs on the M60 at the same time to access the full 16GB, I'm happy to be proven wrong though and I'm sure there are some clever people working on this out there somewhere... :-)

For what it's worth, I don't think it's really a fair comparison, as the M60 is designed for multiple sessions, the K5200 is single session (unless you add it to an RDS / XenApp server...). Rather than try and use this model GPU in a way in which it was not intended, you might be better off looking at an M5000 / M6000 or wait a bit longer now that the Pascal GPUs have been announced and look at a P5000 / P6000 (I believe these will start to become available early October) which are the successors to the K5200 you mention. You can add the Quadro cards to a VM in exactly the same way as the Tesla (M60), but as mentioned, they are a 1-1 mapping, but you can still access them remotely if you wish.

Regards

Ben

谢谢本。
从像您这样的有经验的用户那里获取信息总是好的。
我知道根据您的评论比较不同用途的不同产品是不公平的，至少，我可以期待M60的性能非常接近。
那很好。
我知道P系列尚未上市，也有可能在Nvidia Grid平台上看到Pascal处理器，但我认为我们还有时间。
非常高兴听到K5200也可以在VM中使用，即使它只支持单个会话。
在这种情况下，数据将在远程系统中呈现，如Grid。
对？
所以，我可以使用远程会话来获得GPU的优势。
顺便说一句，您是否知道任何网页或博客一步一步地解释了浮渣类系统的设置。
我知道TurboVNC和其他一些工具的组合可以在远程会话中访问GPU。
对于Xen等商业产品，我是否只需要使用XenApp服务器，或者也可以使用VMVare Horizo​​n等。
无论如何，感谢您的建议和评论。
问候，
--ufuk



以下为原文

Thanks Ben. It is always good to get information from experienced users like you. I know that it is not fair to compare different products that have different purposes but based on your comments, at least, i could expect very close performance from M60. That is nice. I know that P series are not available yet and there might be a possibility to see Pascal processor also in Nvidia Grid platform but i think we have still time for it. That is really nice to hear that K5200 can be also used in VM even if it supports only single session. In this case, the data is rendered in remote system like Grid. Right? So, i could get advantage of GPU using remote session. By the way, are you aware any web page or blog that explains the setup of scum kind of system step by step. I know that the combination of TurboVNC and some other tools could give access to GPU in remote session. For commercial products such as Xen, do i need to use XenApp server only or VMVare Horizon etc. can be also used. Anyway, thanks for your suggestions and comments about it.

Regards,

--ufuk

None


以下为原文

No problem at all, I hope my comments are useful, and if nothing else provide a few other options for you to investigate.

If it were my money, and based on the specific single user scenario you've mentioned, I'd wait till the Pascal Quadro cards are available in a couple of months before upgrading, if this is something you decide to do. Such is the progress with the technology, they will be much more powerful than the K5200. If I wanted to support more than a single user on dedicated VMs, then it has to be an M60 for best performance and capacity. The M60 also has a "Compute" or "Graphics" mode, and can be switched between the 2 depending on workload for flexibility, with that plus the support for multiple users and differing profiles it makes it a much more flexible card than a Quadro, but as said, they have different use cases.

Regarding adding the K5200 into a VM for remote access, this will work, but you may find your resolution is limited as the cards are designed to output through the local Display Connectors (DVI + DP), unlike the Tesla GPUs which don't have any local Display Connectors as they are designed to be used within a server. As a work around, you may be able to use a special "Headless GPU Dongle" that connects to the local Display Connectors, this makes the GPU output at full resolution when accessed remotely. If you purchase a Server for the GPU, make sure your card is on the HCL. Also, you may need a "GPU enablement kit" from that vendor (typically power cables and low profile heat-syncs) and finally you may need upgraded PSUs to support the GPU. Check with the vendor first, or it can get expensive if you have to buy bits afterwards...

Just as a heads up, your GPU desktop, doesn't need to be a VM, it can be a physical PC / Server with a client / agent installed that you access remotely, removing the hypervisor and keeping the performance level. If you use a PC / Workstation for this, then it may be easier to get around any Server HCLs. Obviously using a physical workstation vs virtual workstation has pros and cons too.

There are plenty of remote access solutions out there, with Citrix and VMware being the most common at the moment. I don't really want to get drawn into which is best, as they both have their pros and cons. Hypervisor wise, you're looking at Citrix XenServer, VMware ESXi or KVM.

There's no one guide that will cover every step, as it depends how you want to build the solution. But with all the differing technologies available at the moment, you are spoilt for choice as to what to do. There really are many many different options. However, if you are looking for some information from NVIDIA, here's a good place to start: http://www.nvidia.com/object/grid-enterprise-resources.html#guides

Regards

Ben

我有一台带有K5200卡的Dell T620服务器。
没有GPU支持套件很难运行它并且戴尔不支持K5200（或者没有经过测试配置）但我使用PCI 1x到3x的转换器以及额外的功率输出为GPU获得了额外的功率。
该卡使用两个PIC插槽，但至少它没有任何问题。
所以，我可以开始尝试单用户场景。
无论如何，谢谢你的帮助。
这真的很有用。
问候，
--ufuk



以下为原文

I have a Dell T620 server with K5200 card on it. It was hard to run it without GPU enablement kit and K5200 was not supported (or not tested configuration) by Dell but i got extra power for GPU using an converter from PCI 1x to 3x along with extra power output on it. The card uses two PIC slot but at least it works without any problem. So, i could start to try single user scenario. Anyway, thanks for your help. It was really useful.

Regards,

--ufuk