SOFS / SMB 3 Offers Best VM Resiliency Experience

I have blogged about Virtual Machine Resiliency in Windows 2016 Failover Clustering before in Testing Virtual Machine Compute Resiliency in Windows Server 2016 

Those test and demos were done with block lever storage, CSV on Fibre Channel, iSCSI or shared SAS. Today we’ll look at the experience when you’re running your VMs on a continually available file share on a Scale Out File Server (SOFS). This configuration offers the best possible experience.

Why well, when the cluster node is in Isolated mode this has no impact on the SOFS share as this is a resource external to the Hyper-V cluster. In other words it remains on line. This means that the VMs, even if they have lost their high availability during the time the node is Isolated, they keep running. After all there is nothing wrong with Hyper-V itself. With block level CSV storage you lose access to the storage as that a cluster resource and the node got isolated. That’s why the VMs go into a paused critical state during a transient failure with block level storage but they don’t when you’re using SOFS.

image

The virtual machine compute resiliency feature in action shows you that the VMs service a transient failure without issues. Your services need never know something was up. Even when the transient failure is reoccurring that doesn’t mean it will cause down time. The node will be quarantined and if it come backup the workload will be live migrated away.

image

You can watch a video of this in action here on Vimeo:

The quarantine threshold and duration as well as the resiliency period and can be tweaked to your environment to get the best possible results.

image

SMB 3 for the win! This is yet one more convincing argument to start looking into SOFS and leveraging the capabilities of SMB3. Remember that you can run as SOFS cluster against your existing shared storage to get started if you can get the IOPS/latency you require. But also look into storage spaces, especially storage spaces direct which avoids some of the drawback SANs have in such a scenario. High time for storage vendors to really scale out, implement SMB 3 well and complete and keep the great added value features they already have in their offering. It’s this or becoming yet a bit more irrelevant in todays storage scene in the Microsoft ecosystem.

NUMA Spanning and Virtual NUMA in Hyper-V

When it comes to NUMA Spanning and Virtual NUMA in Hyper-V or anything NUMA related actually in Hyper-V virtualization this is one  subject that too many people don’t know enough about. If they know it they often could be helped by some more in depth information and examples on anything NUMA related in Hyper-V virtualization.

image

Some run everything on the defaults and  never even learn more l they read or find they need to dive in deeper for some needs or use cases. To help out many with some of the confusion or questions they struggled with in regards to Virtual NUMA, NUMA Topology, NUMA Spanning and their relation to static and dynamic memory.image

As I don’t have the time to answer all questions I get in regards to this subject I have written an article on the subject. I’ve published it as a community effort on the StarWind Software blog and you can find it here: A closer look at NUMA Spanning and virtual NUMA settings

I think t complements the information on this subject on TechNet well and it also touches on Windows Server 2016 aspects of this story. I hope you enjoy it!

Setting up Discrete Device Assignment with a GPU

Introduction

Let’s take a look at setting up Discrete Device Assignment with a GPU. Windows Server 2016 introduces Discrete Device Assignment (DDA). This allows a PCI Express connected device, that supports this, to be connected directly through to a virtual machine.

The idea behind this is to gain extra performance. In our case we’ll use one of the four display adapters in our NVIDIA GROD K1 to assign to a VM via DDA. The 3 other can remain for use with RemoteFX. Perhaps we could even leverage DDA for GPUs that do not support RemoteFX to be used directly by a VM, we’ll see.

As we directly assign the hardware to VM we need to install the drivers for that hardware inside of that VM just like you need to do with real hardware.

I refer you to the starting blog of a series on DDA in Windows 2016:

Here you can get a wealth of extra information. My experimentations with this feature relied heavily on these blogs and MSFT provide GitHub script to query a host for DDA capable devices. That was very educational in regards to finding out the PowerShell we needed to get DDA to work! Please see A 1st look at Discrete Device Assignment in Hyper-V to see the output of this script and how we identified that our NVIDIA GRID K1 card was a DDA capable candidate.

Requirements

There are some conditions the host system needs to meet to even be able to use DDA. The host needs to support Access Control services which enables pass through of PCI Express devices in a secure manner. The host also need to support SLAT and Intel VT-d2 or AMD I/O MMU. This is dependent on UEFI, which is not a big issue. All my W2K12R2 cluster nodes & member servers run UEFI already anyway. All in all, these requirements are covered by modern hardware. The hardware you buy today for Windows Server 2012 R2 meets those requirements when you buy decent enterprise grade hardware such as the DELL PowerEdge R730 series. That’s the model I had available to test with. Nothing about these requirements is shocking or unusual.

A PCI express device that is used for DDA cannot be used by the host in any way. You’ll see we actually dismount it form the host. It also cannot be shared amongst VMs. It’s used exclusively by the VM it’s assigned to. As you can imagine this is not a scenario for live migration and VM mobility. This is a major difference between DDA and SR-IOV or virtual fibre channel where live migration is supported in very creative, different ways. Now I’m not saying Microsoft will never be able to combine DDA with live migration, but to the best of my knowledge it’s not available today.

The host requirements are also listed here: https://technet.microsoft.com/en-us/library/mt608570.aspx

  • The processor must have either Intel’s Extended Page Table (EPT) or AMD’s Nested Page Table (NPT).

The chipset must have:

  • Interrupt remapping: Intel’s VT-d with the Interrupt Remapping capability (VT-d2) or any version of AMD I/O Memory Management Unit (I/O MMU).
  • DMA remapping: Intel’s VT-d with Queued Invalidations or any AMD I/O MMU.
  • Access control services (ACS) on PCI Express root ports.
  • The firmware tables must expose the I/O MMU to the Windows hypervisor. Note that this feature might be turned off in the UEFI or BIOS. For instructions, see the hardware documentation or contact your hardware manufacturer.

You get this technology both on premises with Windows Server 2016 as and with virtual machines running Windows Server 2016; Windows 10 (1511 or higher) and Linux distros that support it. It’s also an offering on high end Azure VMs (IAAS). It supports both Generation 1 and generation 2 virtual machines. All be it that generation 2 is X64 bit only, this might be important for certain client VMs. We’ve dumped 32 bit Operating systems over decade ago so to me this is a non-issue.

For this article I used a DELL PowerEdge R730, a NVIIA GRID K1 GPU. Windows Server 2016 TPv4 with CU of March 2016 and Windows 10 Insider Build 14295.

Microsoft supports 2 devices at the moment:

  • GPUs and coprocessors
  • NVMe (Non-Volatile Memory express) SSD controllers

Other devices might work but you’re dependent on the hardware vendor for support. Maybe that’s OK for you, maybe it’s not.

Below I describe the steps to get DDA working. There’s also a rough video out on my Vimeo channel: Discrete Device Assignment with a GPU in Windows 2016 TPv4.

Setting up Discrete Device Assignment with a GPU

Preparing a Hyper-V host with a GPU for Discrete Device Assignment

First of all, you need a Windows Server 2016 Host running Hyper-V. It needs to meet the hardware specifications discussed above, boot form EUFI with VT-d enabled and you need a PCI Express GPU to work with that can be used for discrete device assignment.

It pays to get the most recent GPU driver installed and for our NVIDIA GRID K1 which was 362.13 at the time of writing.

clip_image001

On the host when your installation of the GPU and drivers is OK you’ll see 4 NIVIDIA GRID K1 Display Adapters in device manager.

clip_image002

We create a generation 2 VM for this demo. In case you recuperate a VM that already has a RemoteFX adapter in use, remove it. You want a VM that only has a Microsoft Hyper-V Video Adapter.

clip_image003

In Hyper-V manager I also exclude the NVDIA GRID K1 GPU I’ll configure for DDA from being used by RemoteFX. In this show case that we’ll use the first one.

clip_image005

OK, we’re all set to start with our DDA setup for an NVIDIA GRID K1 GPU!

Assign the PCI Express GPU to the VM

Prepping the GPU and host

As stated above to have a GPU assigned to a VM we must make sure that the host no longer has use of it. We do this by dismounting the display adapter which renders it unavailable to the host. Once that is don we can assign that device to a VM.

Let’s walk through this. Tip: run PoSh or the ISE as an administrator.

We run Get-VMHostAssignableDevice. This return nothing as no devices yet have been made available for DDA.

I now want to find my display adapters

#Grab all the GPUs in the Hyper-V Host

$MyDisplays = Get-PnpDevice | Where-Object {$_.Class -eq “Display”}

$MyDisplays | ft -AutoSize

This returns

clip_image007

As you can see it list all adapters. Let’s limit this to the NVIDIA ones alone.

#We can get all NVIDIA cards in the host by querying for the nvlddmkm

#service which is a NVIDIA kernel mode driver

$MyNVIDIA = Get-PnpDevice | Where-Object {$_.Class -eq “Display”} |

Where-Object {$_.Service -eq “nvlddmkm”}

$MyNVIDIA | ft -AutoSize

clip_image009

If you have multiple type of NVIDIA cared you might also want to filter those out based on the friendly name. In our case with only one GPU this doesn’t filter anything. What we really want to do is excluded any display adapter that has already been dismounted. For that we use the -PresentOnly parameter.

#We actually only need the NVIDIA GRID K1 cards, let’s filter some #more,there might be other NVDIA GPUs.We might already have dismounted #some of those GPU before. For this exercise we want to work with the #ones that are mountedt he paramater -PresentOnly will do just that.

$MyNVidiaGRIDK1 = Get-PnpDevice -PresentOnly| Where-Object {$_.Class -eq “Display”} |

Where-Object {$_.Service -eq “nvlddmkm”} |

Where-Object {$_.FriendlyName -eq “NVIDIA Grid K1”}

$MyNVidiaGRIDK1 | ft -AutoSize

Extra info: When you have already used one of the display adapters for DDA (Status “UnKnown”). Like in the screenshot below.

clip_image011

We can filter out any already unmounted device by using the -PresentOnly parameter. As we could have more NVIDIA adaptors in the host, potentially different models, we’ll filter that out with the FriendlyName so we only get the NVIDIA GRID K1 display adapters.

clip_image013

In the example above you see 3 display adapters as 1 of the 4 on the GPU is already dismounted. The “Unkown” one isn’t returned anymore.

Anyway, when we run

$MyNVidiaGRIDK1 = Get-PnpDevice -PresentOnly| Where-Object {$_.Class -eq “Display”} |

Where-Object {$_.Service -eq “nvlddmkm”} |

Where-Object {$_.FriendlyName -eq “NVIDIA Grid K1”}

$MyNVidiaGRIDK1 | ft -AutoSize

We get an array with the display adapters relevant to us. I’ll use the first (which I excluded form use with RemoteFX). In a zero based array this means I disable that display adapter as follows:

Disable-PnpDevice -InstanceId $MyNVidiaGRIDK1[0].InstanceId -Confirm:$false

Your screen might flicker when you do this. This is actually like disabling it in device manager as you can see when you take a peek at it.clip_image015

When you now run

$MyNVidiaGRIDK1 = Get-PnpDevice -PresentOnly| Where-Object {$_.Class -eq “Display”} |

Where-Object {$_.Service -eq “nvlddmkm”} |

Where-Object {$_.FriendlyName -eq “NVIDIA Grid K1”}

$MyNVidiaGRIDK1 | ft -AutoSize

Again you’ll see

clip_image017

The disabled adapter has error as a status. This is the one we will dismount so that the host no longer has access to it. The array is zero based we grab the data about that display adapter.

#Grab the data (multi string value) for the display adapater

$DataOfGPUToDDismount = Get-PnpDeviceProperty DEVPKEY_Device_LocationPaths -InstanceId $MyNVidiaGRIDK1[0].InstanceId

$DataOfGPUToDDismount | ft -AutoSize

clip_image019

We grab the location path out of that data (it’s the first value, zero based, in the multi string value).

#Grab the location path out of the data (it’s the first value, zero based)

#How do I know: read the MSFT blogs and read the script by MSFT I mentioned earlier.

$locationpath = ($DataOfGPUToDDismount).data[0]

$locationpath | ft -AutoSize

clip_image021

This locationpath is what we need to dismount the display adapter.

#Use this location path to dismount the display adapter

Dismount-VmHostAssignableDevice -locationpath $locationpath -force

Once you dismount a display adapter it becomes available for DDA. When we now run

$MyNVidiaGRIDK1 = Get-PnpDevice -PresentOnly| Where-Object {$_.Class -eq “Display”} |

Where-Object {$_.Service -eq “nvlddmkm”} |

Where-Object {$_.FriendlyName -eq “NVIDIA Grid K1”}

$MyNVidiaGRIDK1 | ft -AutoSize

We get:

clip_image023

As you can see the dismounted display adapter is no longer present in display adapters when filtering with -presentonly. It’s also gone in device manager.

clip_image024

Yes, it’s gone in device manager. There’s only 3 NVIDIA GRID K1 adaptors left. Do note that the device is unmounted and as such unavailable to the host but it is still functional and can be assigned to a VM.That device is still fully functional. The remaining NVIDIA GRID K1 adapters can still be used with RemoteFX for VMs.

It’s not “lost” however. When we adapt our query to find the system devices that have dismounted I the Friendly name we can still get to it (needed to restore the GPU to the host when needed). This means that -PresentOnly for system has a different outcome depending on the class. It’s no longer available in the display class, but it is in the system class.

clip_image026

And we can also see it in System devices node in Device Manager where is labeled as “PCI Express Graphics Processing Unit – Dismounted”.

We now run Get-VMHostAssignableDevice again see that our dismounted adapter has become available to be assigned via DDA.

clip_image029

This means we are ready to assign the display adapter exclusively to our Windows 10 VM.

Assigning a GPU to a VM via DDA

You need to shut down the VM

Change the automatic stop action for the VM to “turn off”

clip_image031

This is mandatory our you can’t assign hardware via DDA. It will throw an error if you forget this.

I also set my VM configuration as described in https://blogs.technet.microsoft.com/virtualization/2015/11/23/discrete-device-assignment-gpus/

I give it up to 4GB of memory as that’s what this NVIDIA model seems to support. According to the blog the GPUs work better (or only work) if you set -GuestControlledCacheTypes to true.

“GPUs tend to work a lot faster if the processor can run in a mode where bits in video memory can be held in the processor’s cache for a while before they are written to memory, waiting for other writes to the same memory. This is called “write-combining.” In general, this isn’t enabled in Hyper-V VMs. If you want your GPU to work, you’ll probably need to enable it”

#Let’s set the memory resources on our generation 2 VM for the GPU

Set-VM RFX-WIN10ENT -GuestControlledCacheTypes $True -LowMemoryMappedIoSpace 2000MB -HighMemoryMappedIoSpace 4000MB

You can query these values with Get-VM RFX-WIN10ENT | fl *

We now assign the display adapter to the VM using that same $locationpath

Add-VMAssignableDevice -LocationPath $locationpath -VMName RFX-WIN10ENT

Boot the VM, login and go to device manager.

clip_image033

We now need to install the device driver for our NVIDIA GRID K1 GPU, basically the one we used on the host.

clip_image035

Once that’s done we can see our NVIDIA GRID K1 in the guest VM. Cool!

clip_image037

You’ll need a restart of the VM in relation to the hardware change. And the result after all that hard work is very nice graphical experience compared to RemoteFX

clip_image039

What you don’t believe it’s using an NVIDIA GPU inside of a VM? Open up perfmon in the guest VM and add counters, you’ll find the NVIDIA GPU one and see you have a GRID K1 in there.

clip_image041

Start some GP intensive process and see those counters rise.

image

Remove a GPU from the VM & return it to the host.

When you no longer need a GPU for DDA to a VM you can reverse the process to remove it from the VM and return it to the host.

Shut down the VM guest OS that’s currently using the NVIDIA GPU graphics adapter.

In an elevated PowerShell prompt or ISE we grab the locationpath for the dismounted display adapter as follows

$DisMountedDevice = Get-PnpDevice -PresentOnly |

Where-Object {$_.Class -eq “System” -AND $_.FriendlyName -like “PCI Express Graphics Processing Unit – Dismounted”}

$DisMountedDevice | ft -AutoSize

clip_image045

We only have one GPU that’s dismounted so that’s easy. When there are more display adapters unmounted this can be a bit more confusing. Some documentation might be in order to make sure you use the correct one.

We then grab the locationpath for this device, which is at location 0 as is an array with one entry (zero based). So in this case we could even leave out the index.

$LocationPathOfDismountedDA = ($DisMountedDevice[0] | Get-PnpDeviceProperty DEVPKEY_Device_LocationPaths).data[0]

$LocationPathOfDismountedDA

clip_image047

Using that locationpath we remove the DDA GPU from the VM

#Remove the display adapter from the VM.

Remove-VMAssignableDevice -LocationPath $LocationPathOfDismountedDA -VMName RFX-WIN10ENT

We now mount the display adapter on the host using that same locationpath

#Mount the display adapter again.

Mount-VmHostAssignableDevice -locationpath $LocationPathOfDismountedDA

We grab the display adapter that’s now back as disabled under device manager of in an “error” status in the display class of the pnpdevices.

#It will now show up in our query for -presentonly NVIDIA GRIDK1 display adapters

#It status will be “Error” (not “Unknown”)

$MyNVidiaGRIDK1 = Get-PnpDevice -PresentOnly| Where-Object {$_.Class -eq “Display”} |

Where-Object {$_.Service -eq “nvlddmkm”} |

Where-Object {$_.FriendlyName -eq “NVIDIA Grid K1”}

$MyNVidiaGRIDK1 | ft -AutoSize

clip_image049

clip_image051

We grab that first entry to enable the display adapter (or do it in device manager)

#Enable the display adapater

Enable-PnpDevice -InstanceId $MyNVidiaGRIDK1[0].InstanceId -Confirm:$false

The GPU is now back and available to the host. When your run you Get-VMHostAssignableDevice it won’t return this display adapter anymore.

We’ve enabled the display adapter and it’s ready for use by the host or RemoteFX again. Finally, we set the memory resources & configuration for the VM back to its defaults before I start it again (PS: these defaults are what the values are on standard VM without ever having any DDA GPU installed. That’s where I got ‘m)

#Let’s set the memory resources on our VM for the GPU to the defaults

Set-VM RFX-WIN10ENT -GuestControlledCacheTypes $False -LowMemoryMappedIoSpace 256MB -HighMemoryMappedIoSpace 512MB

clip_image053

Now tell me all this wasn’t pure fun!

Do you need hard processor affinity in Hyper-V?

Do you need hard processor affinity in Hyper-V? Good question but let’s set the context first. I tend to virtualize workloads that shock some people. Not because they are super huge solutions requiring Petabytes of storage, 48TB of RAM, 256 cores and a million IOPS. Far from. The shock often comes from people who still consider virtualization as something for the lightweight infra services like DHCP, DNS, WSUS, print servers, or web services and websites. Some of these people even tried to virtualize other services like SharePoint ,SQL, Exchange etc. but they did not take into a account that virtualization is not magic and you’ll need to provision adequate resources and design /manage your environment to do so successfully.  So part of them got bitten. They conclude that performance requires physical deployments … and they want to see a material CPU so to speak.

CPU_closeup

When they see virtual machines with 12 tot 16 vCPUs or > 100GB of memory they seem to thinks that even those workloads are bad candidates to virtualize, let alone even bigger ones. That’s not true by definition. As long as you make sure that you know why (cost/benefits/risks) and how to virtualize it can work. You must provision and allocate the required resources. You must also have the right expertise in both virtualization (servers, storage, networking) and the applications involved (SQL, Exchange, 3rd party  products, …)  along with good operational processes.

You can really virtualize a lot when done right. My “virtual first” approach is rule of thumb and exceptions do exist even when I’m calling the shots. However just like people quoting costs, latency, security, lock-in to question the suitability of Public Cloud versus on premises in “subjective” ways, they do so when it comes to virtualization as well. The discussion if often more about organizational issues, control, fear, politics, interests and money. Every hosting provider out there loves virtualization as it’s great for their TCO/ROI. But when it comes to Public Cloud they’re often less convinced. That “datacenter zero” concept isn’t that attractive to them so we see Hybrid and Public Cloud offerings that might not be that good of an idea in some cases but it fits their interests more. Have you noticed that there are no highly automated, optimized data centers anymore, only * clouds? There are valid use cases for hybrid and private clouds but just like with virtualization, maybe we should let go of the personal/business interests, the fear, and false assumptions when advising customers. It all depends.

In this regards I had several discussions now with people about the lack of hard processor affinity with Hyper-V. This makes it unfit for high performance workloads in their opinion. Sure, such cases do exist. These are however, not the majority. As I’ve  been having this discussion rather often in the past months I wrote an article on the subject that I’ve published in collaboration with StarWind Software Need Hard Processor affinity for Hyper-V? The idea is to reach more people and share insights with the community. Full disclosure: I happen to know Anton Kolomyeytsev (CEO, CTO and Chief Architect at StarWind)  professionally as a fellow MVP and I have great respect for his technical expertise, insights and experience. This made me agree to publish some content via their blog. Sharing opinions and ideas with as many people as possible only makes for better technologist everywhere.