Subscribe via e-mail: 
Does server virtualization reduce costs? Part II – Power savings
By Michael Pietroforte | 10 Comments | Permalink | Trackback | Previous | NextThis is my second post in the series about the costs of server virtualization. So, what is it about the “hay argument”: ten horses require ten times more hay than one horse? Do ten servers always consume less energy than one? I don’t think so. How much hay or power is consumed does not depend solely on the number of workers performing a certain task.
I suppose a well-fed sumo wrestler consumes ten times as much food as a ballerina. However, ten ballerinas have a fair chance to floor a sumo wrestler. Power consumption is connected to the amount of work that has to be done. Thus, if a high-end server has to do the work of ten average servers, it obviously needs a lot more power than each of them does.
Of course, most supporters of the power savings claim are aware of this fact. Their point is that servers also consume much power when they are idle. Since a high-end server running ten virtual servers is seldom idle, it will use energy more efficiently. This argument is certainly valid if you assume that the ten average servers are oversized for their tasks and, therefore, often consuming energy without doing valuable work.
However, there are other factors to consider here. Much of the electrical power a server consumes is transferred into heat that has to be abducted. The problem is that heat abduction for an object with a small surface is more difficult than for a big object. This is the reason why the fans in computers are becoming bigger and bigger over the years. I remember quite well that my first computer didn’t even have a fan. The point here is that fans require energy too. The more CPU power that a computer of a certain size has the more energy you will need to get rid of the heat. Obviously, a high-end server has less volume than ten average servers do; this means that you need more energy for cooling this server than for multiple servers having a bigger surface.
Even though the high-end server is well equipped with fans, its CPUs, power supplies, and hard disks will always operate at a higher temperature than those in the ten average servers. This is where the second argument comes in: You will waste more energy for heat (thermal energy) if you run your system at a higher temperature. Imagine a sprinter who has to run one hundred meters in ten seconds. Even if he is well trained he will be sweating a lot. Now imagine ten average persons who have to walk ten meters in ten seconds. Together they also covered one hundred meters, but they won’t sweat. They won’t feel as exhausted as the sprinter will, even if you were to repeat the experiment ten times. However, our runner would most likely be dead afterwards.
I know this metaphor is a bit far-fetched, but I think you understand my point: you can’t always save energy by simply reducing the number of workers. Many factors have to be considered here. If you were to replace ten old servers with one new high-end server, you would probably not measure reduced power consumption afterwards. Perhaps five new average servers without virtualization technology would have been able to do the job as well? Would those five average workers have needed less energy than your sprinter?
If, and by how much, you can reduce in power costs with virtualization technology depends heavily on your environment. I doubt whether a general formula exists, that can calculate that for you. Personally, I do think that one can save power in some cases by virtualizing servers. However, it seems to me that the benefits are often exaggerated by vendors trying to sell high priced high-end servers or virtualization software.
I searched Google for a scientific study that would extend the analysis beyond the typical “idle time argument” but I was not able to find one. This made me even more suspicious. If you are aware of such a study from an independent scientific institution I would be very much interested.
When it comes to the costs of server virtualization, the power savings are outweighed by other factors anyway. This will be the topic of my next post in this series.
Articles in this series:
Well, i’m not sure how big your environment is (a lot bigger than mine).
But i think many of the points go into the direction of extreme enterprises. I recently did a few VMware sales certifications, and there where things like “how many servers does your customers purchase PER MONTH”.
In the small business space were i work, we don’t talk about servers purchases per month – we discuss if we can replace 6 year old machines with new ones. And sometimes even that doesn’t work.
So many of these arguments may be aimed at really, really, big companies. With horribly inefficient processes, where they purchase hundreds of servers per month that do not do anything meaningful – the only reason they exist is The Process.
However, for a typical Small or even Medium Business, where each server purchase is a month long ardous task of trying to get the Powers That Be to agree with spending maybe 50k on new infrastructure, these arguments don’t hold much water.
Virtualization does not save money, as the licenses are almost always as expensive as the servers. Virtualization may allow you to segregate roles, but this is purely to the convencience of IT workers (and would also allow you faster restores, etc.).
Michael,
Intel has a good comparison of energy performance per watt comparison on the newer virtualization chipsets versus the older ones. This shows a 6x performance per watt increase over the older CPU architectures.
http://www.intel.com/performance/server/xeon/ppw.htm?iid=perf_server_lhn+dp_ee
A quad-core PowerEdge 1950 is capable of handling a load > 10 virtual guests and has two 670 watt power supplies. At 208 volt, this would consume about 6.4 amps total ((670 * 2) / 208). Most of the time the system would not be pulling the full 6.4 amps because of the energy efficiency built into the newer systems. It is common sense that the older systems were not as energy efficienct.
Compare this to the power consumption of 10 physical SC1435 servers with single power supplies (30 amps total).
The raw heat in BTU is 1 W = 3.413 BTUs.
So one server with two 600w power supplies would create 4,095 BTUs/hr.
Ten servers with one 600w power supply each would take 20,478 BTUs/hr.
One ton of air conditioning can cool 12,000 Btu’s per hour.
Therefore, ten servers would require about a 2 ton A/C system, whereas a single 1950 would require just a 3rd of one ton.
I do not think the density of the single system outweighs the benefit of reducing BTU output by 80%.
Lukas,
Small Businesses can take advantage of Virtualization technology with little cost. Yes we do work in much larger environments than you do which allows us to use Enterprise class Virtualization technology. Our server purchases have been reduced DRAMATICALLY since we began P2V’ing all of the legacy servers that needed to be refreshed. We analyzed our 70 servers and realized how much harddisk space was wasted not to mention underutilized processors just sitting idle. Virtualizing several servers onto one single host allowed us to manage the disk allocation properly, share the RAM, and allow the processor to manage the guests and host seamlessly. We are now using the whole server to its maximum potential. These servers are repurposed database servers that we already had in service. So we didn’t introduce an expensive server to do the job. These existing servers are hosted in a datacenter and standard office building server room. They are quite efficient at self-cooling and never show heat-related issues.
Your work environment is slow to purchase new servers and budgeting appears hopeless. So use VirtualPC which is FREE! or VMWare Workstation. You can run guest systems on a consumer-grade laptop or desktop computer. This is very useful for Developers and consultants who need to test their projects against a real server or workstation. The guest machine running VirtualPC (on your laptop) behaves seamlessly on the network as if it was in the next room.
I would add that Virtualization does save money, it does allow an end-user to standup a Real server in a matter of minutes instead of waiting for the procurement process.
Lastly, it is not for the convenience of IT staff, it is for the brilliance behind virtualization, how it changed the way we do business, how it opened up so many new doors that were closed just a year ago. Embrace it. I am currently building a 3-server Virtual environment for a Disaster Recovery scenario that is impossible to perform on physical servers. We can restore the data offline from tape and really analyze the data integrity, the process, and build confidence in our backup routine.
Cont: I failed to mention Virtual Server is free as well. Ideally, you want to host on a system that supports Virtualization in the BIOS. This will allow you to secure one decent server, install Virtual Server 2005 R2 and host 3 or 4 virtual guest servers.
Rob, for us (a Microsoft Partner), doing tests and such in a virtual environment is easily possible and often done. However, the cost may still be prohibitive for companies that do not get a lot of free Windows licenses.
Also, it’s not something our customers that just need to run their business (which is not IT related) can profit off.
Software licensing can become more expensive when you use virtualization, e.G. you’ll need more OS licenses and also more application licenses if you for example us Vmotion. Microsoft recently did some adjustments there, but those only apply to VL customers.
Hi Michael,
Great two part post on the costs associated with virtualization. Another thing to note as a risk factor of virtualization is the possibility of the consolidated server becoming the single point of failure for business critical applications. This is where fail safe technology needs to come into play so that an organization isn’t affected by the repercussions of downtime. For those interested, Marathon has an ROI calculator which gives decision makers a quick assessment of virtualization’s potential impact on their organization.
I think power savings depends allot on how you design your VM environment. The use of blade servers over traditional physical servers would help increase your power savings numbers. Another big factor is also what is your virtual machine to physical server ratio. I have talked to admins that only host ten virtual servers per physical host. Their savings would be considerable less then the savings we are seeing with running twenty virtual servers per physical host. Another big factor is if you are allowing your VM solution to automatically manager your environment. Hopefully you are taking advantage of having idle physical servers shutdown at night and powered back on in the morning.
Lukas, I think you hit the nail on the head. Larger organizations tend to buy a new server for every new application. And then, they are surprised how much they can save if they pack more than one app on a physical server. The point is that virtualization isn’t the only way to optimize server utilization.
Joe, I remember quite well when we tried to calculate the power consumptions of all our servers last time. We needed a new air conditioning and so we required reliable numbers. We sent the specs of all our servers to our specialists. They came up with a number, but they were not really sure. Thus, we just measured the power consumptions for all our servers for one week. It turned out that the number we calculated before was about five times too high.
Of course, you can’t just add all the power specs of the power supplies. That is just the maximum they support. You have to know what your servers actually consume, and this depends on the kind of work they have to do. Ten average servers have more power supplies than a high end-server. This is the reason why your calculation shows such a big difference. In my view, all theoretical calculations about this issue are useless. What we need here is empirical data.
This reminds me about the vendor specs for a laptop’s battery life. My batteries usually only last half of the time they say. I guess, I strain my laptop a little more than those marketing guys who wrote those specs.
As to the link to Intel, this says only something about old vs. new processors. Where is the connection to our discussion? Of course, I can run new CPUs in my ten average servers. However, if you replace old servers with new ones you will always measure higher power consumption simply because the new servers are more powerful. I simply made this experience.
Brian, thanks. The single-point-of-failure argument is certainly the most important disadvantage of server virtualization. But since we are talking about costs, software solutions that can remedy this downside are usually not for free.
Jacob, you’re right, power consumption depends on many factors. But as I outlined in the article there are a couple of factors that increase the power costs in a virtual environment. This is the reason why I think it is mere speculation that virtualization always saves power.
Consider the notion of having 10(or 40) “spare” servers waiting for customers of your Service Catalog (typically application development teams) to order and instantly occupy? For those of us in the IT Service Management role, we would not want to lay out capital for excess capacity of physical hosts. You simply cannot do this without Virtualization.
The key to success is consolidation & virtualization – take two of your old dedicated servers and “platespin” them to be your new virtuals. Then decommission the rest – or spin them all to create 100 VM’s from your 10 old hosts – no capital required – cannot possibly consume more energy, and you will have capacity to spare for quite some time.
[...] Does server virtualization reduce costs? Part II – Power savings [...]