My ramblings on the stuff that holds it all together
8 Node ESXi Cluster running 60 Virtual Machines – all Running from a Single 500GBP Physical Server
February 25, 2010Posted by on
The goal of this build is to create a system you can use for VCP and VCDX type study without spending thousands on normal production type hardware (see the slides at the end of this page for more info on why this is useful..) – Techhead and I have a series of joint postings in the pipeline about how to configure the environment and the best hardware to use.
As a bit of a tangent I have been seeing how complex an environment I can get out of a single server (which I have dubbed v.T.A.R.D.I.S: Nano Edition) using virtualized ESXi hosts, the goals were;
- Distributed vSwitch and/or Cisco NX100V
- Cluster with HA/DRS enabled
- Large number of virtual machines
- Single cheap server solution
- No External hardware networking (all internal v/dvSwitch traffic)
The main stumbling block I ran into with the previous build was the performance of the SATA hard disks I was using, SCSI was out of my budget and SATA soon gets bogged down with concurrent requests which makes it slow; so I started to investigate solid state storage (previous posts here).
By keeping the virtual machine configurations light and using thin-provisioning I hoped to squeeze a lot of virtual machines into a single disk, previous findings seem to prove that cheap-er consumer grade SSD’s can support massive amount of IOps when compared to SATA (Eric Sloof has a similar post on this here)
So, I voted with my credit card and purchased one of these from Amazon – it wasn’t “cheap” at c.£200 but it will let me scale my environment bigger than I could previously manage which means less power, cost, CO2 and all the other usual arguments you try to convince yourself that a gadget is REQUIRED.
So the configuration I ended up with is as follows;
|1 x HP ML115G5, 8Gb RAM, 144Gb SATA HDD||c.£300 (see here) but with more RAM|
|1 x 128Gb Kingston 2.5” SSDNow V-Series SSD||c£205|
I installed ESX4U1 classic on the physical hardware then installed 8 x ESXi 4U1 instances as virtual machines inside that ESX installation
This diagram shows the physical server’s network configuration
In order for virtualized ESXi instances to talk to each other you need to update the security setting on the physical host’s vSwitch only as shown below;
This diagram shows the virtual network configuration within each virtualized ESXi VM with vSwitch and dvSwitch config side-side.
I then built a Windows 2008R2 Virtual Machine with vCenter 4 Update 1 as a virtual machine and added all the hosts to it to manage
I clustered all the virtual ESXi instances into a single DRS/HA cluster (turning off admission control as we will be heavily oversubscribing the resources of the cluster and this is just a lab/PoC setup
Cluster Summary – 8 x virtualized ESXi instances – note the heavy RAM oversubscription, this server only has 8Gb of physical RAM – the cluster thinks it has nearly 64Gb
I then built an OpenFiler Virtual Machine and hooked it up to the internal vSwitch so that the virtualized ESXi VMs can access it via iSCSI, it has a virtual disk installed on the SSD presenting a 30Gb VMFS volume over iSCSI to the virtual cluster nodes (and all the iSCSI traffic is essentially in-memory as there is no physical networking for it to traverse.
Each virtualized ESXi node then runs a number of nested virtual machines (VM’s running inside VMs)
In order to get Nested virtual machines to work; you need to enable this setting on each virtualized ESXi host (the nested VM’s themselves don’t need any special configuration)
Once this was done and all my ESXi nodes were running and settled down, I have a script to build out a whole bunch of nested virtual machines to execute on my 8-node cluster. the VM’s aren’t anything special – each has 512Mb allocated to it and won’t actually boot past the BIOS because my goal here is just to simulate a large number of virtual machines and their configuration within vCenter, rather than meet an actual workload – remember this is a single server configuration and you can’t override the laws of physics, there is only really 8Gb or RAM and 4 CPU cores available.
Each of the virtual machines was connected to a dvSwitch for VM traffic – which you can see here in action (the dvUplink is actually a virtual NIC on the ESXi host).
I power up the virtual machines in batches of 10 to avoid swamping the host, but the SSD is holding up very well against the I/O
With all 60 of the nested VMs and virtualized ESXi instances loaded these are the load stats
I left it to idle overnight and these are the performance charts for the physical host; the big spike @15:00 was the scripts running to deploy the 60 virtual machines
Physical memory consumption – still a way to go to get it to 8Gb – who says oversubscription has no use?
So, in conclusion – this shows that you can host a large number of virtual machines for a lab setup, this obviously isn’t of much use in a production environment because as soon as those 60VM’s actually start doing something they will consume real memory and CPU and you will run out of raw resources.
The key to making this usable is the solid state disk – in my previous experiments I found SATA disks just got soaked under load and caused things like access to the VMFS to fail (see this post for more details)
Whilst not a production solution, this sort of setup is ideal for VCP/VCDX study as it allows you to play with all the enterprise level features like dvSwitch and DRS/HA that really need more than just a couple of hosts and VMs to understand how they really work. for example; you can power-off one of the virtual ESXi nodes to simulate a host failure and invoke the HA response, similarly you can disconnect the virtual NIC from the ESXi VM to simulate the host isolation response.
Whilst this post has focused on non-production/lab scenarios it could be used to test VMware patch releases for production services if you are short on hardware and you can quite happily run Update manager in this solution.
If you run this lab at home it’s also very power-efficient and quiet, there are no external cables or switches other than a cross-over cable to a laptop to run the VI Client and administer it; you could comfortably have it in your house without it bothering anyone – and with an SSD there is no hard disk noise under load either
Thin-provisioning also makes good use of an SSD in this situation as this screenshot from a 30Gb virtual VMFS volume shows.
The only thing you won’t be able to play around with seriously in this environment is the new VMware FT feature – it is possible to enable it using the information in this post and learn how to enable/disable but it won’t remain stable and the secondary VM will loose sync with the primary after a while as it doesn’t seem to work very well as a nested VM. If you need to use FT for now you’ll need at least 2 physical FT servers (as shown in the original vTARDIS demo)
If you are wondering how noisy it it at power-up/down TechHead has this video on YouTube showing the scary sounding start-up noise but how quiet it gets once the fan control kicks-in.
ML115 G5 Start-up Noise
Having completed my VCP4 and 3 I’m on the path to my VCDX and next up is the enterprise exam so this lab is going to be key to my study when the vSphere exams are released.