VMware NSX

  • 1.  CPU sizing for NSX-T edges

    Posted Feb 09, 2022 01:47 PM

    Hi,

    I am doing CPU sizing for an NSX-T build and have a question.

    Pre-sales is telling us it's best practice to have one physical core per NSX-T edge vCPU.
    That would mean, with the following CPU:

    Xeon Gold 6354 (18-core @3GHz) x2:

    - 18x2 = 36 physical cores total
    - 2 cores for ESXi/vSAN (36-2 = 34 cores left)
    - assuming Tier0 Large edges (8vCPU): 34/8 = 4 edges max per server.

    Coming from NSX-V, I was under the impression that edges reserve
    1,000MHz per vCPU, which should require 8GHz per Tier0 Large Edge,
    and with the example above, it should allow for
    34 cores x 3 GHz / 8 = 12 edges per server instead of 4.

    It's a big difference, so I was hoping for a second opinion.

    Thanks.



  • 2.  RE: CPU sizing for NSX-T edges

    Broadcom Employee
    Posted Feb 09, 2022 06:09 PM

    vSphere treats Edges as normal VM's, so there is nothing specific about Edge VCPU sizing when you compare it with the overall VCPU ratio. That being said , reservations for Edge CPU is not mandatory and I would recommend only if you have precise requirement and it is tested thoroughly with results showing the clear difference. In the worst case, any fine-tuning in the values will be accepted, however when VM tries to power on if there are not enough resources it will fail to power on - going back to my first point ->vSphere treats Edges as normal VM's. Your primary focus must be mainly on the Edge form factor and type of service requirements for your workloads.



  • 3.  RE: CPU sizing for NSX-T edges

    Posted Feb 10, 2022 02:34 AM

    Thanks for the reply.

    1GHz reservation per vCPU is totally fine with us (we do that with NSX-V).

    I am just having doubts about the 1:1 pCPU:vCPU recommendation ("best practice"?) for NSX-T edges VMware pre-sales is giving us.  It is quite a big difference in hardware and license footprint.

    Thanks.



  • 4.  RE: CPU sizing for NSX-T edges

    Broadcom Employee
    Posted Feb 10, 2022 07:04 AM

    Thanks. I can't comment much on that recommendation, you must certainly cross-check with the respective team to know the reason behind the same considering your design. That being said  I'm also assuming you have a collapsed cluster (Edges and Workloads) and that will make the 1:1 pCPU:vCPU recommendation slightly harder.