The following post was written and scenario tested by my colleagues and friends Keith Keogh and Andrew Breedy of the Limerick, Ireland CCC engineering team. I helped with editing and art. Enjoy.

 

VMWare has added High Availability (HA) for virtual machines (VM’s) with an NVIDIA GRID vGPU shared pass-through device in vSphere 6.5. This feature allows any vGPU VM to automatically restart on another available ESXi host with an identical NVIDIA GRID vGPU profile if the VM’s hosting server has failed. 
To better understand how this feature would work in practice we decided to create some failure scenarios. We tested how HA for vGPU virtual machines performed in the following situations:
1)  Network outage: Network cables get disconnected from the virtual machine host.
2)  Graceful shutdown: A host has a controlled shut down via the ESXi console.
3)  Power outage: A host has an uncontrolled shutdown, i.e. the server loses AC power.
4)  Maintenance mode: A host is put into maintenance mode.

The architecture of the scenario used for testing was laid out as follows:
  

 

Viewed another way, we have 3 hosts, 2 with GPUs, configured in an HA VSAN cluster with a pool of Instant Clones configured in Horizon.

 

1)  For this scenario, we used three Dell vSan Ready Node R730 servers configured per our C7 spec, with the ESXi 6.5 hypervisor installed, clustered and managed through a VMWare vCenter 6.5 appliance.
2)  VMWare VSAN 6.5 was configured and enabled to provide shared storage.
3)  Two of the three servers in the cluster had an NVIDIA M60 GPU card and the appropriate drivers installed in these hosts. The third compute host did not have a GPU card.
4)  16 vGPU virtual machines were placed on one host. The vGPU VM’s were a pool of linked clones created using VMWare Horizon 7.1. The virtual machine Windows 10 master image had the appropriate NVIDIA drivers installed and an NVIDIA GRID vGPU shared PCI pass through device added.
5)  All 4GB of allocated memory was reserved on the VM’s and the M60-1Q profile assigned which has a maximum of 8 vGPU’s per physical GPU. Two vCPU’s were assigned to each VM.
6)  High Availability was enabled on the vSphere cluster.

Note that we intentionally didn’t install an M60 GPU card in the third server of our cluster. This allowed us to test the mechanism that allows the vGPU VM’s to only restart on a host with the correct GPU hardware installed and not attempt to restart the VM’s on a host without a GPU. In all cases this worked flawlessly, the VM’s always restarted on the correct backup host with the M60 GPU installed.

 

Scenario One

To test the first scenario of a network outage we simply unplugged the network cables from the vGPU VM’s host server. The host showed as not responding in the vSphere client and the VM’s showed as disconnected within a minute of unplugging the cables.  After approximately another 30 seconds all the VM’s came back online and restarted normally on the other host in the cluster that contained the M60 GPU card.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Scenario Two

A controlled shutdown down of the host through ESXi worked in a similar manner. The host showed as not responding in the vSphere client and the VM’s showed as disconnected as the ESXi host began shutting down the VSAN services as part of ESXi’s shutdown procedure. Once the host had shut down, the VM’s restarted quite quickly on the backup host. In total, the time from starting the shutdown of the host until the VM’s restarted on the backup host was approximately a minute and a half.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Scenario Three

To simulate a power outage, we simply pulled the power cord of the vGPU VM’s host. After approximately one minute the host showed as not responding and the VM’s showed as disconnected in the vSphere client. Roughly 10 seconds later the vGPU VM’s had all restarted on the back up host.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Scenario Four

Placing the host in maintenance mode worked as expected in that any powered on VMs must first be powered off on that host before completion can occur. Once powered off, the VM’s were then moved to the backup host but not powered on (the ‘Move powered-off and suspended virtual machines to other hosts in the cluster’ option was left ticked when entering maintenance mode). The vGPU-enabled VM’s were moved to the correct host each time we tested this: the host with the M60 GPU card.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

In conclusion, the HA failover of vGPU virtual machines worked flawlessly in all cases we tested. In the context of speed of recovery, we made the following observations with regard to the scenarios tested. The unplanned power outage test recovered the quickest at 1 minute 10 seconds followed by the controlled shutdown, with the network outage taking slightly longer again to recover.

Failure Scenario	Approximate Recovery Time
Network Outage	1 min, 30 sec
Controlled Shutdown	1 min, 30 sec
Power Outage	1 min, 10 sec

In a real world scenario the HA failover for vGPU VM’s would of course require an unused GPU card in a backup server for it to work. In most vGPU environments you would probably want to fill the GPU to its maximum capacity of vGPU based VM’s to gain maximum value from your hardware. In order to provide a backup for a fully populated vGPU host, it would mean having a fully unused server and GPU card lying idle until circumstances required it. It is also possible to split the vGPU VM’s between two GPU hosts, placing, for example 50% of the VM’s on each host. In the event of a failure the VM’s will failover to the other host. In this scenario though, each GPU would be used to only 50% of its capacity.

An unused back-up GPU host may be ok in larger or mission critical deployments but may not be very cost effective for smaller deployments. For smaller deployments it would be recommended to leave a resource buffer on each GPU server so that if there was a host failure, there is room on the GPU enabled server to facilitate these vGPU Enabled VM’s.

I’m publishing this design topic separately for both Hyper-V and vSphere, assuming that the audience will differ between the two. Some of this content will be repeated between posts but arranged to focus on the hypervisor each prospective reader will likely care most about.
Be sure to also check out: Hyper-V Design for NUMA Architecture and Alignment: Link

Non-Uniform Memory Access (NUMA) has been with us for awhile now and was created to overcome the scalability limits of the Symmetric Multi-Processing (SMP) CPU architecture. In SMP, all memory access was tied to a singular shared physical bus. There are obvious limitations in this design, especially with higher CPU counts and increased traffic on the bus.
NUMA was created to limit the number of CPUs tied to a single memory bus and as a result defines what is construed as a NUMA node. A high performing NUMA node is defined by the memory directly attached to a physical CPU socket in turn accessed directly by consuming applications and services. The memory controller is on-die and the memory DIMMs connect directly to the CPU socket.  The total number of logical cores available to a NUMA node is double if Hyperthreading is enabled. Per the example below, I have a dual socket system with 6 physical cores per CPU, including hyperthreading there are 12 logical cores per physical NUMA node, netting a total of 24 logical cores in the system.

NUMA spanning occurs when memory or CPU cores are accessed remote to the NUMA node an application, service or VM is connected. NUMA spanning means your application or service won’t run out of memory or CPU cores within its NUMA node as it can access the remote DIMMs or cores of an adjacent CPU. The problem with NUMA spanning is that your application or service has to cross the lower performing processor interconnects to store its pages in the remote DIMMs or access remote cores. The maximum vCPU value assignable to a VM is dictated by the total number of logical cores available in all NUMA nodes.

What is NUMA alignment?

In the case of virtualization, VMs created on a physical server receive their vCPU entitlements from the physical CPU cores of their hosts (um, obviously). There are many tricks used by hypervisors to oversubscribe CPU and run as many VMs on a host as possible while ensuring good performance. Better utilization of physical resources = better return on investment and use of the company dollar. To design your virtual environment to be NUMA aligned means ensuring that your VMs receive vCPUs and RAM tied to a single physical CPU (pCPU), thus ensuring the memory and pCPU cores they access are directly connected and not accessed via traversal of a CPU interconnect (QPI). As fast as interconnects are these days, they are still not as fast as the connection between a DIMM and pCPU. Ensuing that your virtual environments are NUMA aligned means ensuring the highest possible performance.

Running many smaller 1 or 2 vCPU VMs with lower amounts of RAM, such as virtual desktops, aren’t as big of a concern and we generally see very good oversubscription at high performance levels with less need to span NUMA. In the testing my team does at Dell for VDI architectures, we see somewhere between 9-12 vCPUs per pCPU core depending on the workload and VM profile. For larger server-based VMs, such as those used for RDSH sessions, this ratio is smaller and more impactful because there are fewer overall VMs consuming a greater amount of resources. In short, NUMA alignment is much more important when you have fewer total VMs with larger vCPU and RAM allocations.

To achieve sustained high performance for a dense server workload, it is important to consider:

1. The number of  vCPUs assigned to each VM within a NUMA node to avoid contention.
2. The amount RAM assigned to each VM to ensure that all memory access remains local within a physical NUMA node.

The easiest method to follow is to ensure that your VMs receive vCPUs in an even multiple of the total logical cores available per NUMA node. The following is an example of a balanced configuration based on a dual 6-core, 96GB server used for RDSH. The server hosts 6 high performance VMs that are configured within 2 NUMA node boundaries. Given the number of vCPUs and RAM assigned to each VM, each NUMA node is oversubscribed 2x from a pCPU core perspective (including hyperthreading) and does not exceed the 48GB of RAM available to each pCPU socket. This ensures that NUMA spanning is not required and minimizes potential contention amongst the VMs. Your oversubscription ratio may vary depending on workload or SLAs. For RDSH specifically, the best results are achieved using 1.5x or 2x oversubscription. Here is an example of a NUMA aligned configuration with 4 VMs:

vNUMA in vSphere

In vSphere 5.0 and later, the virtual NUMA topology can be exposed to the guest OS as long as it is hardware version 8 or later. vNUMA is enabled by default if a VM is configured with more than 8 total vCPUs (or 4 vCPU cores + 2 cores per NUMA node) and will match the host where it was powered on until the VM is restarted. This minimum number can be adjusted if desired as shown later.
When the NUMA topology is exposed to a guest VM, one of two things happen via the ESXi NUMA scheduler:

1. If a VM is assigned a vCPU value less than or equal to the number of total logical cores in a physical NUMA node, then this VM will be assigned to a single physical NUMA node and consume memory/ cores local to that node (ideal). This remains true until that VM is configured to access more memory than what is available in the single physical NUMA node, which would result in the VM spanning to the remote NUMA node (less than ideal).
2. If a VM is assigned more vCPUs than cores that exist in a single physical NUMA node, it will be assigned to at least 2 physical NUMA nodes. This will satisfy the resource requirements of the VM but will come with a performance penalty as the VM spans local and remote NUMA nodes (less than ideal).

Some conditions will cause VMs to not be managed by the ESXi NUMA scheduler, such as enabling CPU Hot Add or manual specification of a VM's CPU or memory affinity. This effectively disables vNUMA and carries some obvious inherent risks.
The NUMA scheduler assigns each VM to a “home node” at boot time which is a defined physical NUMA node within the server. ESXi tracks NUMA nodes within the System Resource Allocation Table (SRAT). When calculating a VM's NUMA home node, hyperthreading is ignored. So if your VM is configured with more vCPUs that a single physical CPU has cores, it will be assigned to 2 or more NUMA home nodes. ESXi will ensure VM memory locality by locking a VM to run from the same home node that contains its assigned memory. If system load changes such that a VM’s home node is no longer optimal, ESXi can dynamically migrate a VM to a new home node. Part of this migration could also include a migration of the VM’s memory assignment to prevent spanning NUMA if possible. Important to note that if you set CPU affinity for a VM, this essentially disables the NUMA scheduler for that VM.
There are 2 ways you can verify how many NUMA nodes are present according to what ESXi sees, via esxtop or esxcli. From the console of one of your hosts run esxtop. Once running, press "m" to enable memory stats. Later, to see the NUMA stats as they apply to VMs, from esxtop press "f" to change the fields displayed, select NUMA Stats with "g" and hit any key to exit.


As you can now see, ESXi is aware of 2 physical NUMA nodes in my test system each with 49GB memory directly attached to each node.


The second method is the "easy" way but displays less information overall. Run the following esxcli command which also reveals that this host has 2 NUMA nodes:

esxcli hardware memory get | grep NUMA

Further, just to completely remove any doubt about the number of logical cores available and to which physical NUMA node they belong, consider the following output from esxcli. 12 logical CPUs assigned to each physical node.

esxcli hardware cpu list | grep Node

NUMA Alignment in vSphere

The basic principles of NUMA are the same when dealing with vSphere or Hyper-v however executed slightly differently from a VM perspective. My vSphere cluster shares the same specs of my Windows cluster:

• 2 x Dell PowerEdge R610’s with dual X5670 6-core CPUs and 96GB RAM running vSphere 5.5 U2 in an HA Cluster. 24 total logical hyperthreaded processors as reported by ESXi.

vNUMA can be altered on a VM by VM basis through vCPU assignments or via advanced host settings if global tweaks need to be made. It’s worth noting here that although the vSphere Client will allow you to assign CPU cores as well as “cores per socket” to your VMs, this is generally not required. VMware added this option in previous versions to provide the ability to control the number of sockets seen by the guest OS as a way to get around certain software restrictions and limitations that likely no longer apply. In vSphere a vCPU is a vCPU and will be scheduled according to how it fits into the exposed vNUMA topology. You can specify the cores per socket which will affect the NUMA topology presented to the VM, but also realize that this could have negative impacts as to how well that VM plays within the presented topology. Unless you have a good reason to change this, leave the default value of 1 core per socket when configuring VMs in ESXi. In my test servers a single VM can be configured with up to 24 vCPUs and if it's a Windows server OS, up to 24 sockets, which constitutes the maximum logical configuration of all available cores on this host. Take for example a Server 2012 R2 VM called "X2" configured with 12 vCPUs using the default of 1 core per socket.

Using the Sysinternals Coreinfo tool we can view the virtual NUMA topology presented to the VM, but keep in mind this is not always indicative of the physical NUMA node assignment. With 12 vCPUs assigned, at 1 core per socket (12 sockets), ESXi exposes the VM to 2 vNUMA nodes. Note that this topology is virtual but at 12 vCPUs fits within a single physical NUMA node:

Because the number of vCPUs is >6, ESXi assigns this VM to both NUMA home nodes, shown here in esxtop via the NUMA Home Node (NHN) value. Remember that the NUMA Scheduler ignores the hyperthreaded cores when making this calculation:

Even though at 12 vCPUs this VM fits within the logical boundaries of a single physical NUMA node, this VM is assigned to more vCPUs than the 6 physical cores that exist in a single NUMA node, thus enabling ESXi to reserve the right to schedule this VM's CPU consumption on either NUMA node. If this VM were to be assigned 13 vCPUs, not only would it be assigned to both NUMA home nodes as we see above but it would span NUMA to fulfill it's operational requirements. Per the diagram below you can see that two similarly configured VMs fit within the constraints of each NUMA node from an assigned memory perspective, but because X2 is configured with more vCPUs than total logical cores that exist on one NUMA node, it must cross the QPIs to fully satisfy its entitlement.

Another example of vNUMA in action is the assignment of 24 vCPUs to this VM using an arbitrary 3 cores per socket. vNUMA creates 8 NUMA nodes (sockets) with 3 processors assigned to each.

The other example that would cause a VM to span NUMA would be if it were assigned to more memory than what exists within a single physical NUMA node. For this example I've reconfigured my X2 VM with 6 vCPUs to ensure CPU locality but assigned and reserved more memory than what exists within the NUMA home node to which it is assigned.

For NUMA spanning to actually take place the VM now has to actually consume more memory than what exists in the NUMA home node it was assigned. To do this I used a memory allocation tool and consumed 48GB of active memory. Hopping back over to esxtop we can see exactly the amount of memory being used locally (NLMEM) vs memory spanning to the remote NUMA node (NRMEM).


To illustrate what is happening here, X2 is utilizing memory from its assigned NUMA home node but is also having to access the remote NUMA node to store its memory pages. If the cluster didn't have enough resources available for the configuration of the VM, it would fail to power on.

There are a number of additional tweaks available in the Advanced System Settings for a given host or at the individual VM level. Be extremely careful here and only change something if you have education and purpose.

To allow a VM with fewer than 8 vCPUs to be managed by vNUMA, a configuration parameter entry must be added to its .vmx file. Open the settings of the VM to be configured, under the VM Options tab, expand Advanced and click Edit Configuration beside Configuration Parameters. Look for the entry “numa.vcpu.min” and edit its value, if it’s not there add this entry and specify the number of vCPUs you need in the value. This would be useful in scenarios where you have fewer VMs with larger memory assignments.

The numa.autosize.vcpu.maxPerVirtualNode value assigned from within a VM's configuration parameters, can be used to control the vNUMA topology presented to a VM. By default, based on the physical characteristics of my test servers, the value is 6 vCPUs per node as presented to the VM. This can be adjusted up or down if required.

Key Takeaways

• As a matter of general best practice, assign vCPU values to VMs less than or equal to the number of physical cores on a single NUMA node.
• Designing your virtual infrastructure to work in concert with the physical NUMA boundaries of your servers will net optimal performance and minimize contention.
• A single VM consuming more vCPUs than a single NUMA node contains, will be scheduled across multiple NUMA nodes causing possible CPU contention.
• A single or multiple VMs consuming more RAM than a single NUMA node contains will span NUMA and store a percentage of their pages in the the remote NUMA node causing reduced performance.
• In vSphere vNUMA is enabled by default for VMs with 8+ vCPUs (adjustable) but enabling "hot add CPU" or configuring CPU affinity will essentially disable this feature. Be careful if you do this. 
• When the NUMA Scheduler calculates which NUMA home node to assign a VM to, hyperthreading is ignored. This calculation is based on the physical cores of the CPUs.

Resources:

http://lse.sourceforge.net/numa/faq/
NUMA in ESXi
http://www.vmware.com/pdf/Perf_Best_Practices_vSphere5.5.pdf
https://blogs.vmware.com/vsphere/2013/10/does-corespersocket-affect-performance.html
http://pubs.vmware.com/vsphere-55/topic/com.vmware.ICbase/PDF/vsphere-esxi-vcenter-server-55-resource-management-guide.pdf
https://pubs.vmware.com/vsphere-60/topic/com.vmware.ICbase/PDF/vsphere-esxi-vcenter-server-60-resource-management-guide.pdf

Our latest addition to the wildly popular Dell XC Web-Scale Converged Appliance series powered by Nutanix, is the XC730 for graphics. The XC730 is a 2U node outfitted with 1 or 2 NVIDIA GRID K1 or K2 cards suited for highly scalable and high-performance graphics within VDI deployments.  The platform supports either Hyper-V with RemoteFX or VMware vSphere 6 with NVIDIA GRID vGPU.  Citrix XenDesktop, VMware Horizon or Dell Wyse vWorkspace can be deployed on the platform and configured to run with the vGPU offering. Being a graphics platform, the XC730 currently supports up to 2 x 120w 10, 12, or 14-core Haswell CPUs with 16 or 32GB 2133MHz DIMMs. The platform requires a minimum of 2 SSDs and 4 HDDs with a range of options within each tier. All XC nodes come equipped with a 64GB SATADOM to boot the hypervisor and 128GB total spread across the first 2 SSDs for the Nutanix Home. Don’t know what mix of components to choose or how to size your environment? No problem, we create optimized platforms and validate them to make your life easier.

 

Optimized High Performance Graphics

Our validated platform for the XC730 is a high-end “B7” variant that provides the absolute pinnacle of graphics performance in an HCI offer. 14-core CPUs, 256GB RAM, dual K2’s, 2 x 400GB SSDs and 4 x 1TB HDDs. This is all about graphics so we need more wattage and the XC730 comes equipped with dual 1100w PSUs. iDRAC8 comes standard in the XC Series as well as your choice of 10Gb SFP+ or BaseT Ethernet adapters. I chose this mix of components to provide a very high-end experience to support the maximum number of users allowed by the K2 vGPU profiles. Watch this space for additional optimized platforms designed to remove the guess of sizing graphics in VDI. Even when GPUs are present, the basic laws of VDI exist which state: thou shalt exhaust CPU first. Dual 14-core parts will ensure that doesn’t happen. Make sure to check out our Appliance Architectures at the bottom of this post for more detailed info.

 

 

NVIDIA GRID vGPU

The cream of the crop in server virtualized graphics comes courtesy of NVIDIA with the K1 and K2 Kepler-based boards. The K2 has 4x the CUDA cores of the K1 board with a slightly lower core clock and less total memory. The K2 is the board you want for fewer higher end users. The K1 is appropriate for a higher overall number of lower end graphical users. vGPU is where the magic happens as graphics are hardware-accelerated to the virtual environment running within vSphere.  Each desktop VM within the vGPU environment runs a native set of the NVIDIA drivers and addresses a real slice of the server grade GPU. The graphic below from NVIDIA portrays this quite well.

 

(image courtesy of NVIDIA)

 

The magic of vGPU happens within the profiles that ultimately get assigned. Both K1 and K2 in conjunction with vGPU have a preset number of profiles that ultimately control the amount of graphics memory assigned to each desktop along with the total number of desktops that can be supported by each card. The profile pattern is K2xxx for K2 and K1xxx for K1, the smaller index identifying a great user density. The K280Q and K180Q are basically pass-through profiles where an entire physical GPU is passed through to a desktop VM. You’ll notice how the densities change per card and per GPU depending on how these profiles are assigned. Lower performance = greater densities with up to 64 total users possible on an XC730 with dual K1 cards. In the case of the K2, the maximum number of users possible on a single XC730 with dual cards is 32.

 

So how does it fit?

When designing a solution using the XC730, all of the normal Nutanix and XC rules apply including a minimum of 3 nodes in a cluster and a CVM running on every node. Since the XC730 is a graphics appliance it is recommended to have additional XC630s in the architecture to run the VDI infrastructure management VMs as well as for the users not requiring high-performance graphics. Nothing will stop you from running mgmt infa on your XC730 nodes but financial practicality will probably dictate that you use XC630s for this purpose. Scale of either is unlimited.

 

 

It is entirely acceptable (and expected) to mix XC730 and XC630 nodes within the same NDFS cluster. The boundaries you will need to draw will be around your vSphere HA clusters separating each function into a discrete cluster up to 64 nodes, as depicted below. This allows each discrete layer to scale independently while providing dedicated N+1 HA for each function. Due to the nature of graphics virtualization, vMotion is not currently supported neither is automated HA VM restarts when GPUs are attached. HA in this particular scenario would be cold should a node in the cluster fail. Using pooled desktops and mobilizing user data, this should only equate to a minor inconvenience worst case. 

 

 

As is the case with any Nutanix deployment, very large NDFS clusters can be created with multiple hypervisor pods created within a singular namespace. Below is an example depiction of a 30K user deployment within a single site. Each compute pod is composed of a theoretical maximum number of users within a VDI farm, serviced by a dedicated pair of mgmt nodes for each farm. Mgmt is separated into a discrete cluster for all farms and compute is separated per the HA maximum. This is a complicated architecture but demonstrates the capabilities of NDFS when designing for a very large scale use case.

 

This is just the tip of the iceberg! For more information on the XC series architecture, platforms, Nutanix, test results of our XC730 platform and more, please see our latest Appliance Architectures which include the XC730:

XC Series for Citrix

XC Series for VMware

XC Series for vWorkspace

Following our launch of the new 13G-based XC series platform, I present our product architectures for the VDI-specific use cases. Of the platforms available, this use case is focused on the extremely powerful 1U XC630 with Haswell CPUs and 2133MHz RAM. We offer these appliances on both Server 2012 R2 Hyper-V and vSphere 5.5 U2 with Citrix XenDesktop, VMware Horizon View, or Dell vWorkspace.  All platform architectures have been optimized, configured for best performance and documented.

Platforms

We have three platforms to choose from optimized around cost and performance, all being ultimately flexible should specific parameters need to change. The A5 model is the most cost effective leveraging 8-core CPUs, 256GB RAM 2 x 200GB SSDs for performance and 4 x 1TB HDDs for capacity.  For POCs, small deployments or light application virtualization, this platform is well suited. The B5 model steps up the performance by adding four cores per socket, increasing the RAM density to 384GB and doubling the performance tier to 2 x 400GB SSDs. This platform will provide the best bang for the buck on medium density deployments of light or medium level workloads. The B7 is the top of the line offering 16-core CPUs and a higher capacity tier of 6 x 1TB HDDs. For deployments requiring maximum density of knowledge or power user workloads, this is the platform for you.

At 1U with dual CPUs, 24 DIMM slots and 10 drive bays…loads of potential and flexibility!

Solution Architecture

Utilizing 3 platform hardware configurations, we are offering 3 VDI solutions on 2 hypervisors. Lots of flexibility and many options. 3 node cluster minimum is required with every node containing a Nutanix Controller VM (CVM) to handle all IO. The SATADOM is present for boot responsibilities to host the hypervisor as well as initial setup of the Nutanix Home area. The SSDs and NL SAS disks are passed through directly to each CVM which straddle the hypervisor and hardware. Every CVM contributes its directly-attached disks to the storage pool which is stretched across all nodes in the Nutanix Distributed File System (NDFS) cluster. NDFS is not dependent on the hypervisor for HA or clustering. Hyper-V cluster storage pools will present SMB version 3 to the hosts and vSphere clusters will be presented with NFS. Containers can be created within the storage pool to logically separate VMs based on function. These containers also provide isolation of configurable storage characteristics in the form of dedupe and compression. In other words, you can enable compression on your management VMs within their dedicated container, but not on your VDI desktops, also within their own container. The namespace is presented to the cluster in the form of \\NDFS_Cluster_name\container_name.
The first solution I’ll cover is Dell’s Wyse vWorkspace which supports either 2012 R2 Hyper-V or vSphere 5.5. For small deployments or POCs we offer this solution in a “floating mgmt” architecture which combines the vWorkspace infrastructure management roles and VDI desktops or shared session VMs. vWorkspace and Hyper-V enables a special technology for non-persistent/ shared image desktops called Hyper-V Catalyst which includes 2 components: HyperCache and HyperDeploy. Hyper-V Catalyst provides some incredible performance boosts and requires that the vWorkspace infrastructure components communicate directly with the hyper-V hypervisor. This also means that vWorkspace does not require SCVMM to perform provisioning tasks for non-persistent desktops!

• HyperCache – Provides virtual desktop performance enhancement by caching parent VHDs in host RAM. Read requests are satisfied from cache including requests from all subsequent child VMs.
• HyperDeploy – Provides instant cloning of parent VHDs massively diminishing virtual desktop pool deployment times.

You’ll notice the HyperCache components included on the Hyper-V architectures below. 3 to 6 hosts in the floating management model, depicted below with management, desktops and RDSH VMs logically separated only from a storage container perspective by function. The recommendation of 3-7 RDSH VMs is based our work optimizing around NUMA boundaries. I’ll dive deeper into that in an upcoming post. The B7 platform is used in the architectures below.

Above ~1000 users we recommend the traditional distributed management architecture to enable more predictable scaling and performance of both the compute and management hosts. The same basic architecture is the same and scales to the full extent supported by the hypervisor, is this case Hyper-V which supports up to 64 hosts. NDFS does not have a scaling limitation so several hypervisor clusters can be built within a single contiguous NDFS namespace. Our recommendation is to then build independent Failover Clusters for compute and management discretely so they can scale up or out independently.
The architecture below depicts a B7 build on Hyper-V applicable to Citrix XenDesktop or Wyse vWorkspace.


This architecture is relatively similar for Wyse vWorkspace or VMware Horizon View on vSphere 5.5 U2 but fewer total compute hosts per HA cluster, 32 total. For vWorkspace, Hyper-V Catalyst is not present in this scenario so vCenter is required to perform desktop provisioning tasks.

For the storage containers, the best practice of less is more still stands. If you don’t need a particular feature don’t enable it, as it will consume additional resources. Deduplication is always recommended on the performance tier since the primary OpLog lives on SSD and will always benefit. Dedupe or compression on the capacity tier is not recommended, of course you absolutely need it. And if you do prepare to increase each CVM RAM allocation to 32GB.

Container	Purpose	Replication Factor	Perf Tier Deduplication	Capacity Tier Deduplication	Compression
Ds_compute	Desktop VMs	2	Enabled	Disabled	Disabled
Ds_mgmt	Mgmt Infra VMs	2	Enabled	Disabled	Disabled
Ds_rdsh	RDSH Server VMs	2	Enabled	Disabled	Disabled

Network Architecture

As a hyperconverged appliance, the network architecture leverages the converged model. A pair of 10Gb NICs minimum in  each node handle all traffic for the hypervisor, guests and storage operations between CVMs. Remember that the storage of all VMs is kept local to the host to which the VM resides, so the only traffic that will traverse the network is LAN and replication. There is no need to isolate storage protocol traffic when using Nutanix. 
Hyper-V and vSphere are functionally similar. For Hyper-V there are 2 vSwitches per host, 1 external that aggregates all of the services of the host management OS as well as the vNICs for the connected VMs. The 10Gb NICs are connected to a LBFO team configured in Dynamic Mode. The CVM alone connects to a private internal vSwitch so it can communicate with the hypervisor.

In vSphere it’s the same story but with the concept of Port Groups and vMotion.

We have tested the various configurations per our standard processes and documented the performance results which can be found in the link below. These docs will be updated as we validate additional configurations.

Product Architectures for 13G XC launch:

Resources:

About Wyse vWorkspace HyperCache
About Wyse vWorkspace HyperDeploy
SJbWUYl4LRCwFNcatHuG

If you’re unfamiliar, PVS (Citrix Provisioning Server) is a vDisk deployment mechanism available for use within a XenDesktop or XenApp environment that uses streaming for image delivery. Shared read-only vDisks are streamed to virtual or physical targets in which users can access random pooled or static desktop sessions. Random desktops are reset to a pristine state between logoffs while users requiring static desktops have their changes persisted within a Personal vDisk pinned to their own desktop VM. Any changes that occur within the duration of a user session are captured in a write cache. This is where the performance demanding write IOs occur and where PVS offers a great deal of flexibility as to where those writes can occur. Write cache destination options are defined via PVS vDisk access modes which can dramatically change the performance characteristics of your VDI deployment. While PVS does add a degree of complexity to the overall architecture, since its own infrastructure is required, it is worth considering since it can reduce the amount of physical computing horsepower required for your VDI desktop hosts. The following diagram illustrates the relationship of PVS to Machine Creation Services (MCS) in the larger architectural context of XenDesktop. Keep in mind also that PVS is frequently used to deploy XenApp servers as well.

PVS 7.1 supports the following write cache destination options (from Link):

• Cache on device hard drive - Write cache can exist as a file in NTFS format, located on the target-device’s hard drive. This write cache option frees up the Provisioning Server since it does not have to process write requests and does not have the finite limitation of RAM.
• Cache on device hard drive persisted (experimental phase only) - The same as Cache on device hard drive, except cache persists. At this time, this write cache method is an experimental feature only, and is only supported for NT6.1 or later (Windows 7 and Windows 2008 R2 and later).
• Cache in device RAM - Write cache can exist as a temporary file in the target device’s RAM. This provides the fastest method of disk access since memory access is always faster than disk access.
• Cache in device RAM with overflow on hard disk - When RAM is zero, the target device write cache is only written to the local disk. When RAM is not zero, the target device write cache is written to RAM first.
• Cache on a server - Write cache can exist as a temporary file on a Provisioning Server. In this configuration, all writes are handled by the Provisioning Server, which can increase disk IO and network traffic.
• Cache on server persistent - This cache option allows for the saving of changes between reboots. Using this option, after rebooting, a target device is able to retrieve changes made from previous sessions that differ from the read only vDisk image.

Many of these were available in previous versions of PVS, including cache to RAM, but what makes v7.1 more interesting is the ability to cache to RAM with the ability to overflow to HDD. This provides the best of both worlds: extreme RAM-based IO performance without the risk since you can now overflow to HDD if the RAM cache fills. Previously you had to be very careful to ensure your RAM cache didn’t fill completely as that could result in catastrophe. Granted, if the need to overflow does occur, affected user VMs will be at the mercy of your available HDD performance capabilities, but this is still better than the alternative (BSOD).

Results

Even when caching directly to HDD, PVS shows lower IOPS/ user numbers than MCS does on the same hardware. We decided to take things a step further by testing a number of different caching options. We ran tests on both Hyper-V and ESXi using our standard 3 user VM profiles against LoginVSI’s low, medium, high workloads. For reference, below are the standard user VM profiles we use in all Dell Wyse Datacenter enterprise solutions:

Profile Name	Number of vCPUs per Virtual Desktop	Nominal RAM (GB) per Virtual Desktop	Use Case
Standard	1	2	Task Worker
Enhanced	2	3	Knowledge Worker
Professional	2	4	Power User

We tested three write caching options across all user and workload types: cache on device HDD, RAM + Overflow (256MB) and RAM + Overflow (512MB). Doubling the amount of RAM cache on more intensive workloads paid off big netting a near host IOPS reduction to 0. That’s almost 100% of user generated IO absorbed completely by RAM. We didn’t capture the IOPS generated in RAM here using PVS, but as the fastest medium available in the server and from previous work done with other in-RAM technologies, I can tell you that 1600MHz RAM is capable of tens of thousands of IOPS, per host. We also tested thin vs thick provisioning using our high end profile when caching to HDD just for grins. Ironically, thin provisioning outperformed thick for ESXi, the opposite proved true for Hyper-V. To achieve these impressive IOPS number on ESXi it is important to enable intermediate buffering (see links at the bottom). I’ve highlighted the more impressive RAM + overflow results in red below. Note: IOPS per user below indicates IOPS generation as observed at the disk layer of the compute host. This does not mean these sessions generated close to no IOPS.

Hyper-visor	PVS Cache Type	Workload	Density	Avg CPU %	Avg Mem Usage GB	Avg IOPS/User	Avg Net KBps/User
ESXi	Device HDD only	Standard	170	95%	1.2	5	109
ESXi	256MB RAM + Overflow	Standard	170	76%	1.5	0.4	113
ESXi	512MB RAM + Overflow	Standard	170	77%	1.5	0.3	124
ESXi	Device HDD only	Enhanced	110	86%	2.1	8	275
ESXi	256MB RAM + Overflow	Enhanced	110	72%	2.2	1.2	284
ESXi	512MB RAM + Overflow	Enhanced	110	73%	2.2	0.2	286
ESXi	HDD only, thin provisioned	Professional	90	75%	2.5	9.1	250
ESXi	HDD only thick provisioned	Professional	90	79%	2.6	11.7	272
ESXi	256MB RAM + Overflow	Professional	90	61%	2.6	1.9	255
ESXi	512MB RAM + Overflow	Professional	90	64%	2.7	0.3	272

For Hyper-V we observed a similar story and did not enabled intermediate buffering at the recommendation of Citrix. This is important! Citrix strongly recommends to not use intermediate buffering on Hyper-V as it degrades performance. Most other numbers are well inline with the ESXi results, save for the cache to HDD numbers being slightly higher.

Hyper-visor	PVS Cache Type	Workload	Density	Avg CPU %	Avg Mem Usage GB	Avg IOPS/User	Avg Net KBps/User
Hyper-V	Device HDD only	Standard	170	92%	1.3	5.2	121
Hyper-V	256MB RAM + Overflow	Standard	170	78%	1.5	0.3	104
Hyper-V	512MB RAM + Overflow	Standard	170	78%	1.5	0.2	110
Hyper-V	Device HDD only	Enhanced	110	85%	1.7	9.3	323
Hyper-V	256MB RAM + Overflow	Enhanced	110	80%	2	0.8	275
Hyper-V	512MB RAM + Overflow	Enhanced	110	81%	2.1	0.4	273
Hyper-V	HDD only, thin provisioned	Professional	90	80%	2.2	12.3	306
Hyper-V	HDD only thick provisioned	Professional	90	80%	2.2	10.5	308
Hyper-V	256MB RAM + Overflow	Professional	90	80%	2.5	2.0	294
Hyper-V	512MB RAM + Overflow	Professional	90	79%	2.7	1.4	294

Implications

So what does it all mean? If you’re already a PVS customer this is a no brainer, upgrade to v7.1 and turn on “cache in device RAM with overflow to hard disk” now. Your storage subsystems will thank you. The benefits are clear in both ESXi and Hyper-V alike. If you’re deploying XenDesktop soon and debating MCS vs PVS, this is a very strong mark in the “pro” column for PVS. The fact of life in VDI is that we always run out of CPU first, but that doesn’t mean we get to ignore or undersize for IO performance as that’s important too. Enabling RAM to absorb the vast majority of user write cache IO allows us to stretch our HDD subsystems even further, since their burdens are diminished. Cut your local disk costs by 2/3 or stretch those shared arrays 2 or 3x. PVS cache in RAM + overflow allows you to design your storage around capacity requirements with less need to overprovision spindles just to meet IO demands (resulting in wasted capacity).
References:
DWD Enterprise Reference Architecture
http://support.citrix.com/proddocs/topic/provisioning-7/pvs-technology-overview-write-cache-intro.html
When to Enable Intermediate Buffering for Local Hard Drive Cache

Dell Desktop Virtualization Solutions/ Cloud Client Computing is proud to announce support for the Dell VRTX on DVS Enterprise for Citrix XenDesktop.  Dell VRTX is the new consolidated platform that combines blade servers, storage and networking into a single 5U chassis that we are enabling for the Remote Office/ Branch Office (ROBO) use case. There was tremendous interest in this platform at both the Citrix Synergy and VMworld shows this year!

Initially this solution offering will be available running on vSphere with Hyper-V following later, ultimately following a similar architecture to what I did in the Server 2012 RDS solution. The solution can be configured in either 2 or 4 M620 blades, Shared Tier 1 (cluster in a box), with 15 or 25 local 2.5” SAS disks. 10 x 15K SAS disks are assigned to 2 blades for Tier 1 or 20 disks for 4 blades. Tier 2 is handled via 5 x 900GB 10K SAS. A high-performance shared PERC is behind the scenes accelerating IO and enabling all disk volumes to be shared amongst all hosts. 
Targeted sizing for this solution is 250-500 XenDesktop (MCS) users with HA options available for networking. Since all blades will be clustered, all VDI sessions and mgmt components will float across all blades to provide redundancy. The internal 8-port 1Gb switch can connect up to existing infrastructure or a Force10 switch can be added ToR. Otherwise there is nothing externally required for the VRTX VDI solution, other than the end user connection device. We of course recommend Dell Wyse Cloud Clients to suit those needs. 

For the price point this one is tough to beat. Everything you need for up to 500 users in a single box! Pretty cool.
Check out the Reference Architecture for Dell VRTX and Citrix XenDesktop: Link

Like all good stewards of all things virtual I need stay current on the very foundations of everything we do:  hypervisors. So this post contains my notes on upgrading to the new vSphere 5.5 build from 5.1. This isn’t meant to be an exhaustive step-by-step 5.5 upgrade guide, as that’s already been done, and done very well (see my resources section at the bottom for links). This is purely my experience upgrading with a few interesting call outs along the way that I felt worth writing down should anyone else encounter them.

The basic upgrade sequence goes like this, depending on which of these components you have in play:

1. vCloud components
2. View server components
3. vCenter (and vSphere Clients)
4. SRM/ VR/ vCOPS/ VDP/ VSA
5. ESXi hosts
6. VM Tools (and virtual hardware –see my warning in the clients section)
7. vShield components
8. View agent

The environment I’m upgrading consists of the following:

• 2 x Dell PE R610 hosts running ESXi 5.1 on 1GB SD
• 400GB of local SAS
• 3TB Equallogic iSCSI storage
• vCenter 5.1 VM (Windows)

 

vCenter 5.5

As was the case in 5.1, there are still 2 ways to go with regard to vCenter: Windows-based or vCenter Server Appliance (vCSA). The Windows option is fully featured and capable of supporting the scaling maximums as well as all published configurations. The vCSA is more limited but in 5.5, a bit less so. Most get excited about the vCSA because it’s a Linux appliance (no Windows license), it uses vPostgres (no SQL license) and is dead simple to set up via OVF deployment. The vCSA can use only Oracle externally, MS SQL Server support appears to have been removed. The scaling maximums of the vCSA has increased to 100 hosts/ 3000 VMs with the embedded database which is a great improvement over the previous version. There are a few things that still require the Windows version however, namely vSphere Update Manager and vCenter Linked Mode.

While I did deploy the vCSA, additionally, my first step was upgrading my existing Windows vCenter instance. The Simple Install method should perform a scripted install of the 4 main components. This method didn’t work completely for me as I had to install the inventory and vCenter services manually. This is also the dialog from which you would install the VUM service on your Windows vCenter server.

I did receive an error about VPXD failing to install after the vCenter Server installation ran for the first time. A quick reboot cleared this up. With vCenter upgraded, the vSphere Client also needs to be upgraded anywhere you plan to access the environment using that tool. VMware is making it loud and clear that the preferred method to manage vSphere moving forward is the web client.

vCenter can support up to 10,000 active VMs now on a single instance, which is a massive improvement. If you plan to possibly power on more than 2000 VMs simultaneously, make sure to increase the number of ephemeral ports during the port configuration dialog of the vCenter setup.

Alternatively, the vCSA is very easy to deploy and configure with some minor tweaks necessary to connect to AD authentication. Keep in mind that the maximum number of VMs supported on the vCSA with the embedded DB is only 3000. To get the full 10,000 out of the vCSA you will have to use an Oracle DB instance externally. The vCSA is configured via the web browser and is connected to by the web and vSphere Clients the same way as its Windows counterpart.

If you fail to connect to the vCSA through the web client and receive a “Failed to connect to VMware Lookup Service…” error like this:

…from the admin tab in the vCSA, select yes to enable the certificate regeneration option and restart the appliance.

 

Upgrading ESXi hosts:

The easiest way to do this is starting with hosts in a HA cluster attached to shared storage, as many will have in their production environments. With vCenter upgraded, move all VMs to one host, upgrade the other, rinse, repeat, until all hosts are upgraded. Zero downtime. For your lab environments, if you don’t have luxury of shared storage, 2 x vCenter servers can be used to make this easier as I’ll explain later. If you don’t have shared storage in your production environment and want to try that method, do so at your own risk.

There are two ways to go to upgrade your ESXi hosts: local ISO (scripted or not) or vSphere Update Manager. I used both methods, one on each, to update my hosts.

ISO Method

The ISO method simply requires that you attach the 5.5 ISO to a device capable of booting your server: either USB or Virtual Media via IPMI (iDRAC). Boot, walk through the steps, upgrade. If you’ve attached your ISO to Virtual Media, you can specify within vCenter to boot your host directly to the virtual CD to make this a little easier. Boot Options for your ESXi hosts are buried on the Processors page for each host in the web or vSphere Client.

 

 

VUM method:

Using VUM is an excellent option, especially if you already have this installed on your vCenter server.

• In VUM console page, enter “admin view”
• Download patches and upgrades list
• Go to ESXi images and import 5.5 ISO
• Create baseline during import operation

• Scan hosts for non-compliance, attach baseline group to the host needing upgrade
• Click Remediate, walk through the screens and choose when to perform the operation

The host will go into maintenance mode, disconnect from vCenter and sit at 22% completion for 30 minutes, or longer depending on your hardware, while everything happens behind the scenes. Feel free to watch the action in KVM or IPMI.

When everything is finished and all is well, the host will exit maintenance mode and vCenter will report a successful remediation.

Pros/ Cons:

• The ISO method is very straight-forward and likely how you built your ESXi host to begin with. Boot to media, upgrade datastores, upgrade host. Speed will vary by media used and server hardware, whether connected to USB directly or ISO via Virtual Media.
• This method requires a bit more hand-holding. IPMI, virtual media, making choices, ejecting media at the right time…nothing earth shattering but not light touch like VUM..
• If you already have VUM installed on your vCenter server and are familiar with its operation, then upgrading it to 5.5 should be fairly painless. The process is also entirely hands-off, you press go and the host gets upgraded magically in the background.
• The down side to this is that the vSphere ISO is stored on and the update procedure is processed from your vCenter server. This could add time delays to load everything from vCenter to your physical ESXi hosts, depending on your infrastructure.
• This method is also only possible using the Windows version of vCenter and is one of the few remaining required uses of the vSphere Client.

No Shared Storage?

Upgrading hosts with no shared storage is a bit more trouble but still totally doable without having to manually SCP VM files between hosts. The key is using 2 vCenter instances in which vCSA works great as a second for this purpose. Simply transfer your host ownership between vCenter instances. As long as both hosts are “owned” by the same vCenter instance, any changes to inventory will be recognized and committed. Any VMs reported as orphaned should also be resolved this way. You can transfer vCenter ownership back and forth any number of times without negatively impacting your ESXi hosts.  Just keep all hosts on one or the other! The screenshot below shows 2 hosts owned by the Windows vCenter (left) and the vCSA on the right showing those same hosts disconnected. No VMs were negatively impacted doing this.

The reason this is necessary is because VM migrations are controlled and handled by vCenter, cold migrations included which is all you’ll be able to do here. Power off all VMs, migrate all VMs off one host, fire up vCenter on the other host, transfer host ownership, move the final vCenter VM off then upgrade that host. Not elegant but it will work and hopefully save some pain.

 

vSphere Clients

The beloved vSphere Client is now deprecated in this release and comes with a heavy push towards exclusive use of the Flash-based web client. My advice, start getting very familiar with the web client as this is where the future is heading, like it or not. Any new features enabled in vCenter 5.5 will only be accessible via the web client.

Here you can see this clearly called out on the, now, deprecated vSphere Client login page:

 

WARNING - A special mention needs to be made about virtual hardware version 10. If you upgrade your VMs to the latest hardware version, you will LOSE the ability to edit their settings in the vSphere Client. Yep, one step closer to complete obsolesce. If you’re not yet ready to give up using the vSphere Client, you may want to hold off upgrading the virtual hardware for a bit.

vSphere Web Client

The web client is not terribly different from the fat client but its layout and operational methods will take some getting used to. Everything you need is in there, it may just take a minute to find it. The recent tasks pane is also off to the right now, which I really like.

The familiar Hosts and Clusters view:

Some things just plain look different, most notably the vSwitch configuration. Also the configuration items you’ve grown used to being in certain property menus are now spread out and stored in different places. Again, not bad just…different.

I also see no Solutions and Applications section in the web client, only vApps. So things like the Dell Equallogic Virtual Storage Manager would have to be accessed via the old vSphere Client.

Client Integration Plugin

To access features like VM consoles, file transfers to datastores and OVF deployments via the web client, the 153MB Client Integration plugin must be installed. Attempting to use any features that require this should prompt you for the client install. One of the places it can also be found is by right-clicking while browsing within a datastore.

The installer will launch and require you to close IE before continuing.

 

VM consoles now appear in additional browser tabs which will take some getting used to. Full screen mode looks very much like a Windows RDP session which I like and appreciate.

Product Feature Request

This is a very simple request (plea) to VMware to please combine the Hosts and Clusters view with the VMs and Templates view. I have never seen any value in having these separated. Datacenters, hosts, VMs, folders and templates should all be visible and manageable from a single pane. There should be only 3 management sections separating the primary vSphere inventories: Compute (combined view), storage and network. Clean and simple.

Resources:

vSphere 5.5 upgrade order (KB): Link

vSphere 5.5 Configuration Maximiums: Link

Awesome and extensive vSphere 5.5 guide: Link

Resolution for VCA FQDN error: Link