I’ve been analysing our CPU utilisation figures and I’m impressed by how well Hyper-V is running and upset at how much hardware we’ve used.
We’re a hosting company. We have no idea what our physical requirements will be for our private cloud. A customer might come in wanting a low end virtual web server. Or they might come in and want some highly available a working floating point number crunching VM. I don’t have crystal ball or a Ouija board so how am I to know?
I looked at the figures this week and what I’m seeing is that we are barely touching the CPU resources in our hosts. We started off guessing what our future requirements would be. We put in dual quad core CPU’s and hoped for the best. What I’m seeing today is full hosts (RAM-wise) that barely touch their available CPU resources. I have a recently purchased and filled single CPU host that averages around 25% CPU utilisation.
Our unpredictability makes us different to most who are implementing virtualisation for the first time. Most people already have existing physical servers that they plan to migrate to their new virtualisation platform. They can run some monitoring (either done by a consultant for the specific task) or use something like Operations Manager with the additional VMM reports. In this scenario if a server uses 50% of it’s existing single quad core CPU then it’s going to use 25% or thereabouts of the resources on a dual quad core CPU host (assuming similar generation of CPU).
What about low end requirements? How many of those sorts of VM’s can you get on Hyper-V? Windows Server 2008 R2 Hyper-V is very scalable. You can have up to 64 logical processors (physical CPU cores) in a host (Enterprise and Datacenter editions). Each logical processor can support up to a maximum of 8 virtual processors (that’s the CPU assigned to a VM). A standalone host can have up to 384 virtual processors. A clustered node can have up to 64 running virtual processors. Make sure your storage and RAM can handle the associated loads on their resources first.
Now I have some empirical data that I can make semi-informed decisions on (no crystal ball, remember!). Our next purchases will feature the Nehalem processors which increase the load capacity over our current hosts. We will increase the RAM capacity per host, thus increasing the amount of VM’s per host. I think I’ll be going single CPU. That will reduce power, purchase and SPLA licensing costs. Between VMM and OpsMgr, I’ll know if that needs to increase.