Horizon 8 Enterprise - Sizing - Best CPU- Nutanix

[PetrosG]

Hello Community, 

I just need some input on your VDI infrastructure - I will share mine:

We use 8 Node Nutanix cluster on Horizon 8 2206 with Esxi 7.0 Update 3

CPU - Intel Xeon Gold 6226 2.7 GHZ (2)  --> 12 core Total 24

Nutanix

768 GB RAM

1GB Tesla M10 GPU Profile

Almost everything I can think of is optimized to use Hardware Acceleration.

Most users are running 3 cores and 8 GB of RAM

Login Time is below 30 seconds with Fslogix, UEM/DEM/ and App volumes.

Windows 10 with O365 and Adobe and McAfee Move with Agent

VDI Performance is acceptable.

My issue is My ESXi hosts get to the 90 % CPU Utilization.

I have 330 VDIs in total. Remember this is on 8 ESXI Hosts.

I see Intel and Cisco UCS recommendations where they are able to run 150 or more on a similar infrastructure as far as CPU and Memory.

I am thinking of replacing Nutanix altogether and use my All Flash FS7300 Storage Array that seem to have low microsecond latency on my Server cluster. This will require me to replace my Nutanix Appliances and I am thinking about Gen 4 Gold  or Platinum CPU with 32 Cores for a total of 64 cores with Nvidia A16 GPU and faster memory.

could your share your VDI infrastructure that you are happy with or have similar issues as me?

Also, do any of you use the Intel Optane Persistent Memory for VDI?

Thoughts on going with Gold or Platinum CPU? 

Thoughts on benefits of DDR4 to DDR 5 Memory or Adding Optane Persistent Memory?

Any input is greatly appreciated in advance.

[Yoshimotto]

Hi,

What is your container storage policy on your nutanix ?

Have you got some compression & déduplication ? What is the sizing of yours CVM (vCPU & RAM) ?

As you know some enabled options need vCPU & RAM on the CVM.

Rgds

[PetrosG]

Yoshimotto, 

Thank you for your feedback and questions:

1. container storage policy in Nutanix : Compression and cache are NOT Checked when I checked on Prisim.

2. CVM have  -  1 Socket and 12 cores

    Memory 32.GB

Let me know what you think please.

[Yoshimotto · Accepted solution]

Hi,

For me you haven't enougth CPU.

330 VDI x 3vCore by VDI = 990 vCore.

990 vCore / 8 Hosts =247.5 vCore / Host

Your hosts have 24 Cores & the recommandation for VDI is 1 or 2 vCore / pCore.

So, 48/96 vCore with the HT. You have a pCore/vCore ratio > 10. 

[PetrosG]

Yoshimotto, Thank you so much for your insight.

I am planning on replacing the 8 node cluster with 5 nodes 

Limit my users to 2 cores (it will be a better and faster processor than currently used)

Use faster memory.

Upgrade to A16 from M10

2x 32 core per host = 64 cores per host 64*5 = 320 total cores cluster

330vdi*2= 660 vCore

660vCores / 5 = 132 vCore / Host

with 64 cores per host, I should maintain less than 2.1  pCore/vCore ratio.

I may be forced to start with 4 node and add one soon after

[PetrosG]

Admin, 

can you help me correct the Best Solution? Yoshimotto Provided the best solution and somehow, I marked mine as the solution by mistake. so sorry Yoshimotto. hope it gets corrected asap.

[Lieven]

Which vGPU profile are you using on your VDIs?

Are you using persistent or non-persistent VDIs?

[PetrosG]

using nonpersistent and for GPU, I am using the M10 1 GB profile.

[Lieven]

Hi Petros,

I have a very similar setup.

6 Nutanix nodes with (unfortunately) varying CPUs. The varying CPUs are due to the fact that there are more and more VDIs being used and therefore we needed to extend the capacity. We initially started with 3 nodes, later on added another node and two years ago another 2:

• 3 nodes with Xeon(R) E5-2650 v4 CPUs (24 cores, 2.2 GHz), 382 GB RAM and Tesla M10 GPUs
• 1 nodes with Xeon(R) Gold 6226 CPUs (24 cores, 2.7 GHz), 382 GB RAM and Tesla M10 GPUs
• 2 nodes with Xeon(R) Gold 6226R CPUs (32 cores, 2,9 GHz), 382 GB RAM and Tesla M10 GPUs

CVMs have 12 vCPUS and 36 GB RAM

Hypervisor: ESXi 7 U3

Horizon 8 2303, FSLogix, DEM, App Volumes

Non-persistent VDIs:

• Pool1:
  • Win 10
  • 4 CPU
  • 8 GB RAM
  • GPU profile: grid_m10-1b
  • Office 365 (including NS Teams), Adobe, 7-zip, Notepad++, other business software
  • SentinelOne Antivirus
• Pool2: 
  • Win 10
  • 6 CPU
  • 12 GB RAM
  • GPU profile: grid_m10-1b
  • Office 365 (including NS Teams), Adobe, 7-zip, Notepad++, other business software
  • AutoCAD
  • Adobe Creative Cloud
  • SentinelOne Antivirus

Logon times = +/- 45 seconds during peak hours, +/- 30 seconds during off-peak hours

We are running +/- 110 VDIs concurrently (+ 15 spare, powered on VDIs). 

90% of the VDIs is from Pool1, 10% of the VDIs is from Pool2

Due to the varying CPUs the load of the nodes is very different (varying from 50 to 90%). On the nodes with the best CPUs we can run +/- 28 VDIs.

The bottleneck for us currently is CPU load on our oldest Nutanix nodes (> 5 years old). 

Cluster-wide disk latency = +/- 1ms (with peaks to +/- 2 ms)

Cluster-wide IOPS = +/- 10000 (with peaks to +/- 18000)

We are thinking of replacing the entire infrastructure with 6 new (identical) nodes:

• NX-3155GN-G8 Nutanix nodes
• 2x Intel Xeon-Platinum 8360Y processors (2.4 GHz/ 36-core)
• 256 GB RAM
• NVIDIA A16 GPU

This config should allow us to run +/- 140 concurrent VDI sessions without any problem with some room for growth.

Regards,

Lieven

[PetrosG]

HI Lieven, 

Thanks for sharing .. those 3 E5 CPU are probably your bottle neck. The thing that I do not link about Nutanix is that the CVM have a large CPU and memory requirements.

I started my purchase process for the 4 Dell 760 servers with 2 A16 and 32 core 2.8 Platinum CPU (latest Gen) DDR5 and I am going to initially try it with 2 cores per user (since the CPU is faster along with memory) and with total 716 GHz available (currently 519 GHz on 8 nodes) with out the CVM overhead, apple to apple, I should be able to run my 300 VDIs.

I will move Test Pools and other odd persistent VDI to a reduced current Nutanix cluster.

on your side, consider the DDR5 with Intel New Generation CPUs and you can pick either a gold or platinum CPU. you already have the 36 cores - you will have 384 cores. with a a much lower Physical core to Virtual core ratio, you should be golden.

Thanks