TCO / ROI

– The vTrooper Report –

This is a continuation of the Capacity Conundrum, if you missed the first part start here.

$ per Compute VM

So let’s cut to the chase.  In the case of the compute tiles of our Quad we have a price per vCPU and $ per GB of RAM to settle. Keeping our example 2U server in play we could expect to spend approximately $15,000 for a 2U fully loaded with 4GB DIMMs.   Well unfortunately a small part of that 15K  is consumed in I/O cards and maintenance which needs to be pulled out to get the compute number.  For our argument we will use $10K for the compute system without the I/O cards and maint. costs;  This is the CAPEX we will offset in our $/per values.

vCOMPUTE – FIREPOWAH!

We know how a CPU works right? Move process into memory , execute CPU cycles, churn, churn more, back to the I/O guys, rinse and repeat.  Basically, this is where the hardware container happens in our data centers.   I say container because it’s easy to show it as a box; It’s hard to define what it will always be in physical form.  1U, 2U, 4U, half blade, Full Blade, appliance , PC ; you name it, it is probably in some one’s ‘datacenter’.  The lowest common denominator I have been able to settle on for a common form factor is Cores per Ram.  Grouping per socket fits because you are measuring the type of memory that is close to the CPU socket.  The NUMA architectures of AMD and Intel with memory controllers on-board and transports to the memory DIMMs without access through the I/O  controllers (eg. Northbridge) help define the grouping.

TECHNOTE:  Every core has associated memory banks it will use and every container(physical server) has a series of sockets that it controls.   A hypervisor has a limit to how well it can control the associated memory space to the nearest vCPU.   Generally the hypervisor will always schedule available vCPU’s from the same socket and swap the corresponding memory for those processes to the memory banks of the corresponding socket.  It does this is for efficiencies of the x86 architectures.  It can move the vm to another socket and readdress the memory but it has a ‘cost’ associated with such a move.  Path of least resistance is to stay in the same socket.

If you create a 4 vCPU VM and run it on a 1 core  processor it gets bogged down.   If you have the same VM on a two socket Quad Core (8 Cores)  the four cores utilized by the VM are likely to be on socket 1 or socket 2 .  The cost of splitting the vCPU between the two physical sockets by the scheduler is greater than running the vCPU in the same socket.    AMD delivered this earlier than Intel and sustained higher levels of virtualization consolidation “Per Host” than similar class systems of Intel could provide through the Northbridge.   Core i7 is a new game for Intel and the results of Nehalem show the improvements.

For more indepth information here is a good read:  CPU Scheduler in VMware ESX

We have a host of $10K  CapX charge that has two sockets at a 4/45GB Socket Ratio with approx $5k spend in each socket.  Looking at our Hardware invoice the CPU Cores are about 25% of the cost of a socket so we can assume that our per socket cost is broken down into 25% Core and 75% Memory.   So our Socket Ratio yields a $1250 cost for 4 cores and $3750 for 45GB of memory:

Per Core CapX = $312.50;  Per GB RAM CapX = $83.33

That gives us a bare metal cost without a hypervisor charge on top, but we need a hypervisor to get a VM running.  Adding in the ESX cost for a per socket license of ESX Enterprise Plus (worst case) you can add $3500 each socket.

ESX Lic. Cost per socket CapX = $3500

Raw burn rate of the host would be $8500 per socket if we never loaded a VM on the Host.  Well, we did it for a reason, so let’s get our money back. If we target the standard allocation for this host (4/45GB socket ratio) we get our target VM count of 16 per socket(1 vCPU/2.8 GB RAM).  Also, keep in mind that we broke the socket cost down by 25%  to CPU and 75% to Memory so we will keep that  same  split here.  If we don’t do the split, then any VM that is deployed to the socket will bear the same cost regardless of its size.

ESX Lic. Cost per VM= $218.75   ( 3500 / 16 )

-Or-

Split by the 25/75 % we did previously for the cost of the CPU and Memory and you get a little different calculation.

3500 * .25 = 875 / 16  = $55 AND     3500 *.75 = 2625 / 45  = $58

per vCPU=$55

per vMEM=$58

Adding it up with our target ratios in tow we get the burn rate of the $ per Compute on a VM basis.

($312/4  = 4:1 ratio) + (83*2.8) + {(55*1) + (58 * 2.8) } = $530

Or Summarized: (vCPU = $78)+(vMEM = $233)+(Hyper$=219)=(vCOMPUTE = $530)

Assuming 8760 Hours (1 year) this VM would cost $.06/hr in vCOMPUTE.

Lets apply that to some other VM systems and see if it sticks.  If we plan for the following VM deployment on our socket:

The costs would spit out as such:

Or slice it up into a per hour number:

So based on this analysis some of my VM’s probably only cost $.05 per hour for vCompute.  Interesting. What is more interesting is the fact that the memory cost associated with a VM scales more accurately to the consumption.  You can have as much memory you like for your new 4 and 8 GB aspirations; (eg. memory leaks) you just need to pay for it accordingly.

Too bad that only pays for the top part of my total cost model.  That said, the benefit here is that this model can span across hypervisors and any market hypervisor can be split up to show the cost of a VM consumed on a Xen , KVM, VirtualIron, Parallels’, or Hyper-V infrastructure.

I will be working on a few powershell scripts and excel calculators that one can use to make this model more repeatable. At the very least, it is a model that I will use to consider CapacityIQ and third party products like the offering from VKernel; and the output they measure.  Especially if they consume additional costs on a per socket basis.  Which I can now calculate as Overhead.

Alas there is more to consider, stay tuned for Part III – “the I/O that binds”

–The vTrooper Report –

In an effort to gel up an internal billing and allocation model (GaaS – Goughing as a Service) I’ve been struggling with the concept of cost per vm.  I was asking a simple question in the twittersphere about that idea and it turned into a discussion and well…got out of hand.  I apologize for that, as this is a better format to explain. (Special thanks to @asweemer for a dumping ground)

If I had a Nickel for each VM…

At VMWorld 2009 there was a presentation in the keynote that showed the price of a vm hosted with Terramark that was $.05 per hour.  I thought  wow.   A nickel per hour.  If I had a nickel per vm/hour; How much would I have available to spend on coffee?

Then I thought wait.  I have VM’s. How much do they really cost me per hour?  Well the answer is … it depends.  Old servers with high power consumption and low density versus a new system with Intel 5500’s and packed in blades have different burn rates visible to different systems(power, cooling, depreciation).  I haven’t found a great model to break those units down to my satisfaction yet.  I need another way.

As a general practice I create in my mind some maxims that I follow in the creation of a VM.

1. S  – 1vCPU, 1GB Ram, 1GB Net, 10GB Disk
2. M -2vCPU, 2GB Ram, 1GB Net, 20GB Disk
3. L – 4vCPU, 4GB Ram, 2GB Net, 40GB Disk

Seems simple enough, but it doesn’ t really generate a cost model on a consistant basis.   Hardware continues to change and each VM that consumes resources does it at different rates and times of the day.   A VM that isn’t doing anything isn’t really ‘consuming’ anything, right?  I thought I would try break it down further by creating a 4 quadrant block with two macro categories:  Compute (CPU and Memory) and I/O (Network and Disk)

Each resource area could increase\decrease for a reason without changing the size of the original maxim it was created under.  This allows for small variations of size without having a customer yell that their bill when up by $2 this month.

The Measureable Unit

Use a unit of measure to identify the four quadrants:   vCPU : vMEM : vNET : vDISK  or C:M:N:D  .   Then overlay the VM creation to count up the units.  This way the growth of a ‘VM’ during its lifecycle can be adequately allocated back into the proper IT metric.  Using the VM creation maxims up above this may be:

1. S  – 1:1:1:1
2. M -2:2:1:2
3. L – 4:4:2:4

This isn’t perfect but it at least allows for the average cpu cost to be allocated seperately from a memory, network, and disk cost.  Afterall, you don’t get to upgrade all four parts of the quadrant in the same fiscal year usually.  This also allows a way to trend an average of your cost rate per unit over a period of months and years to see which cost areas are improving.  It is an interesting metric for the business and IT.  Win-Win in my book.  Even if no-one internally ever has to pay the values back (Showback).  It also helps police which VM is consuming too much of a specific value which would skew the numbers if you simply took the cost of the esx hosts and divide by the number of VM’s.

Apples , Oranges, Lemons, and Grapes = Frutti Results

So you have a unit of measure and a type of system to match the measurement up towards over a period of time.  Here’s where the fruit cart and the horse get hooked up.

This is all very complex, why can’t I just buy the same server I have purchased for the last 5 years?

Sorry Kids. They don’t build’em like they used to.  But in todays market, the UCS system from Cisco has a new buzz to the original players of IBM, HP, and Dell.   How do you sort any of that out among the offerings, and how do you select the right platform for your new ESX System?     By the Socket !   Every system of the x86 family has them from both the Intel and AMD families.  And now that you have to pay for your hypervisor and additional tools (Capacity IQ, AppSpeed,  Nexus1000v) per socket  it matters more.  I need to squeeze the value out of those sockets.

Still staying in the upper half of the Quad;  lets measure cores and RAM as a ratio assuming dual rank 4GB Dimms and measure them to some of the standard 2 socket servers.

Standard Intel x5450

2 Socket – 4 Core – 16 Dimms (8 per socket) produces 4 cores/ 32 GB Ram

Standard Intel Nehalem  x5500

2 Socket – 4 Core – 18 Dimms (9 per socket)  produces 4 cores/ 45 GB Ram

Cisco UCS extention  on x5500

2 Socket – 4 Core – 48 Dimms (24 per socket)   produces 4 cores/ 96 GB ram

What this shows is that for every license of ESX consumed in the environment there are different amounts of memory available for a VM to use.  The approach by the UCS system allows for a much higher allowance of memory to a VM at the same licensing cost.   Sure you could buy 4 way servers and claim that the 256 GB of RAM gives the VM more allowance but in reality the vm will have ratios of contention to the vCPU and Memory within each of the 4 sockets. You can change the size of the container by moving to a 4 way,  but it won’t change the value of the ratio  for that container in regards to the cores and memory.

CPU Contention

The idea of CPU contention is becoming more visible to most administrators of virtualized environments because the desire to pack the vm’s onto a host is so strong.  If I can get 10 VM’s on a host for $5000 then getting 25 VM’s on the same host is lowering my cost per vm.  It could also be cheating your customers of the performance they paid. Especially if you have multiple vCPU’s assigned to those 25 VM’s.    This is where the ratio of VM per host becomes obsolete and vCPU/core  makes more sense.

Using the example containers above you can generate an expected number of VM’s per socket.  There is no reason to do a 1:1 ratio of cores to VM because the point of virtualization is to run more with less.  I think a good ratio to start with is 4:1 for a production VM and 16:1 for a VDI implementation:

Standard Intel x5450  -  (4 /32 GB SocketRatio)   yields 16  VM’s with a 1 vCPU/ 2GB ram configuration per socket

Standard Intel Nehalem  x5500  -  (4 /45GB SocketRatio) yields 16  VM’s with a 1 vCPU/ 2.8GB ram configuration per socket

Cisco UCS  -   (4 /96GB SocketRatio) yields 16  VM’s with a 1 vCPU/ 6GB ram configuration per socket

You can always adjust your actual deployment if these ratios don’t match up for your environment.   The expected deployment number helps determine how large the pizza slices are for the team.  Not how many slices each of them consume.  In these configurations you can see where the density of the RAM per socket (SocketRatio) of the UCS allows for much larger VM configurations before overcommitment. A nice fit for the new 64bit installations. These expected numbers of VM per socket help determine what the burn rate of a C:M:N:D value is for the CapX  spend you made.

BurnRate

To fully understand how much a VM costs, one has to look at what was spent in the CapX of the host and agree on the measuring stick to measure the C:M:N:D value of the created VM.  If a series of hosts are in service from different families and are at different parts of lifecycle there may have to be some averaging.  The SocketRatio of Cores/RAM is a consistent way to measure systems from different form factors and families and levelset the expected allocation of VM’s.  The expected allocation of VM’s for a host helps determine what density ratio is desired for vCPU:vMEM.

This is the end of Part 1 –  In Part Deux I will take a deeper dive into the Compute and I/O areas and assign a more detail cost per VM model.