Virtualization, Cloud, Infrastructure and all that stuff in-between

My ramblings on the stuff that holds it all together

Category Archives: Run ESX on a cheap Desktop PC

8 Node ESXi Cluster running 60 Virtual Machines – all Running from a Single 500GBP Physical Server

 

I am currently presenting a follow-up to my previous vTARDIS session for the London VMware Users Group where I demonstrated a 2-node ESX cluster on cheap PC-grade hardware (ML-115g5).

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

image

This diagram shows the physical server’s network configuration

image

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;

image

This diagram shows the virtual network configuration within each virtualized ESXi VM with vSwitch and dvSwitch config side-side.

image

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

image

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

image

image

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.

image

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)

image

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).

image

image

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

image

 

 

 

 

 

 

 

 

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

image

Disk Latency

image

Physical memory consumption – still a way to go to get it to 8Gb – who says oversubscription has no use? 🙂

image image

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.

image

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.

New Home Lab Design

 

I have had a lab/test setup at home for over 15 years now, it’s proven invaluable to keep my skills up to date and help me with study towards the various certifications I’ve had to pass for work, plus I’m a geek at heart and I love this stuff 🙂

over the years it’s grown from a BNC based 10mbit LAN running Netware 3/Win 3.x, through Netware 4/NT4, Slackware Linux and all variants of Windows 200x/RedHat.

Around 2000 I started to make heavy use of VMware Workstation to reduce the amount of hardware I had (8 PCs in various states of disrepair to 2 or 3 homebrew PCs) in latter years there has been an array of cheap server kit on eBay and last time we moved house I consolidated all the ageing hardware into a bargain eBay find – a single Compaq ML570G1 (Quad CPU/12Gb RAM and an external HDD array) which served fine until I realised just how much our home electricity bills were becoming!

Yes, that's the beer fridge in front of the rack :) hot & cold aisles, mmm 

Note the best practice location of my suburban data centre, beer-fridge providing hot-hot aisle heating, pressure washer conveniently located to provide fine-mist fire suppression; oh and plenty of polystyrene packing to stop me accidentally nudging things with my car. 🙂

I’ve been using a pair of HP D530 SFF desktops to run ESX 3.5 for the last year and they have performed excellently (links here here and here) but I need more power and the ability to run 64 bit VMs (D530’s are 32-bit only) I also need to start work on vSphere which unfortunately doesn’t look like it will run on a D530.

So I  a acquired a 2nd-hand ML110 G4 and added 8Gb RAM – this has served as my vSphere test lab to-date, but I now want to add a 2nd vSphere node and use DRS/HA etc. (looks like no FT for me unfortunately though) – Techhead put me onto a deal that Servers Plus are currently running so I now have 2 x ML110 servers 🙂 they are also doing quad-core AMD boxes for even less money here – see Techhead for details of how to get free delivery here

image

In the past my labs have grown rather organically as I’ve acquired hardware or components have failed; being as this time round I’ve had to spend a fair bit of my own money buying items I thought it would be a good idea to design it properly from the outset 🙂

The design goals are:

  • ESX 3.5 cluster with DRS/HA to support VM 3.5 work
  • vSphere DRS/HA cluster to support future work and more advanced beta testing
  • Ability to run 64-bit VMs (for Exchange 2007)
  • Windows 2008 domain services
  • Use clustering to allow individual physical hosts to be rebuilt temporarily for things like Hyper-V or P2V/V2P testing
  • Support a separate WAN DMZ and my wireless network
  • Support VLAN tagging
  • Adopt best-practice for VLAN isolation for vMotion, Storage etc. as far as practical
  • VMware Update manager for testing
  • keep ESX 3/4 clusters seperate
  • Resource pool for “production” home services – MP3/photo library etc.
  • Resource pool for test/lab services (Windows/Linux VMs etc.)
  • iSCSI SAN (OpenFiler as a VM) to allow clustering, and have all VMs run over iSCSI.

The design challenges are:

  • this has to live in my garage rack
  • I need to limit the overall number of hosts to the bare minimum
  • budget is very limited
  • make heavy re-use of existing hardware
  • Cheap Netgear switch with only basic VLAN support and no budget to buy a decent Cisco.

Luckily I’m looking to start from scratch in terms of my VM-estate (30+) most of them are test machines or something that I want to build separately, data has been archived off so I can start with a clean slate.

The 1st pass at my design for the ESX 3.5 cluster looks like the following

 image

I had some problems with the iSCSI VLAN, and after several days of head scratching I figured out why; in my network the various VLANs aren’t routable (my switch doesn’t do Layer 3 routing). For iSCSI to work the service console needs to be accessible from the iSCSI VKernel port. In my case I resolved this by adding an extra service console on the iSCSI VLAN to get round this problem and discovery worked fine immediately

image image

image

I also need to make sure the Netgear switch had the relevant ports set to T (Tag egress mode) for the VLAN mapping to work – there isn’t much documentation on this on the web but this is how you get it to work.

image

The vSwitch configuration looks like the following – note these boxes only have a single GbE NIC, so all traffic passes over them – not ideal but performance is acceptable.

imageimage

iSCSI SAN – OpenFiler

In this instance I have implemented 2 OpenFiler VMs, one on each D530 machine, each presenting a single 200Gb LUN which is mapped to both hosts

Techhead has a good step-by-step how to setup an OpenFiler here that you should check out if you want to know how to setup the volumes etc.

I made sure I set the target name in Openfiler to match the LUN and filer name so it’s not too confusing in the iSCSI setup – as shown below;

if it helps my target naming convention was vm-filer-X-lun-X which means I can have multiple filers, presenting multiple targets with a sensible naming convention – the target name is only visible within iSCSI communications but does need to be unique if you will be integrating with real-world stuff.

image

Storage Adapters view from an ESX host – it doesn’t know the iSCSI target is a VM that it is running 🙂

image

Because I have a non routed L3 network my storage is all hidden in the 103 VLAN, to administer my OpenFiler I have to use a browser in a VM connected to the storage VLAN, I did play around with multi-homing my OpenFilers but didn’t have much success getting iSCSI to play nicely, it’s not too much of a pain to do it this way and I’m sure my storage is isolated to a specific VLAN.

The 3.5 cluster will run my general VMs like Windows domain controllers, file servers and my SSL VPN, they will vMotion between the nodes perfectly. HA won’t really work as the back-end storage for the VM’s live inside an OpenFiler, which is a VM – but it suits my needs and storage vMotion makes online maintenance possible with some advanced planning.

Performance from VM’d OpenFilers has been pretty good and I’m planning to run as many as possible of my VMs on iSCSI – the vSphere cluster running on the ML110’s will likley use the OpenFilers as their SAN storage.

This is the CPU chart from one of the D530 nodes in the last 32hrs whilst I’ve been doing some serious storage vMotion between the OpenFiler VM’s it hosts.

image

image

image

That’s it for now, I’m going to build out the vSphere side of the lab shortly on the ML110’s and will post what I can (subject to NDA, although GA looks to be close)

Slow vMotion..

 

Note to remember, don’t forget to check the duplex settings on NICs handling your vMotion traffic.

My updated clustered ESX test lab is progressing (more posts on that in the next week or so)… and I’m kind of limited in that I only have an old 24-port 100Mb Cisco hub for the networking at the moment.

vMotion warns about the switch speed as a possible issue.

image

I had my Service Console/ vMotion NICit forced to 100/full and when I 1st tried it vMotion took 2hrs to get to 10%, I changed it to auto-negotiate whilst the task was running and it completed without breaking the vMotion task ain a couple of seconds, dropped only 1 ping to the VM I moved.

Cool, it’s not production or doing a lot of workload but useful to know despite the warning it will work even if you’ve only got an old hub for your networking, and worth remembering that Duplex mis-matches can literally add hours and days onto network transfers.

Performance on a cheap ESX PC

 

I thought I’d post some performance graphs from my cheap HP D530 ESX server using the Virtual Centre console (which incidentally, is good for getting this info quickly and simply).

Screenshot of the UI for querying performance stats.

image

View of currently running VMs – a mix of Windows 2003/2008 VMs

image

Current Overall ESX Host statistics (with a clone from template going on)

image

As I noted elsewhere on my blog it has 4Gb RAM and a single 2.8GHz HT CPU – and with this VM load it gives an average CPU load of 25-30%. Almost all of these VM’s are idling but all respond in good time to network access/TS etc- not bad at all for a desktop PC!

CPU usage for the last 24 hours

image

The big spike around 22:00 was when I cloned up a whole load more VM’s – seems to have upset the stats so need to try and have a look at that..

It’s also interesting to note that I added 4 Windows 2003 VM’s last night but that hasn’t actually increased the overall CPU average – ESX must be quite efficient at time-slicing all those idle VMs.

I had 3-4 “deploy from template..” operations going on at the same time and it really bogged down the performance of the VM’s (usable, but only just..) but it is just a single SATA disk drive so I can live with that.

Deploying 1 VM at a time had little or no impact – slight CPU spike to ~50% as you’ll see to the far right of the chart as I kicked off another one just now.

When i get time I’m going to drop some jobs into the VM’s that will tax the virtual CPUs a bit more and compare results – maybe some Folding@Home activity Mmmmm that would definitley tax it.