Green Telecom’s Tony Chan speaks to Rackspace Hosting chairman, Graham Weston, on the company’s approach to green practices, what’s makes a green data centre and what’s the real significance of virtualisation.
Green Telecom: So what is Rackspace Hosting’s approach to ‘green’ practices?
Graham Weston: Two three years ago, we actually switched to AMD processors because they drew less energy. We intentionally, by one decision, we went with AMD. It saved on our power bill, our carbon footprint, it helps us be more efficient and be more productive in the data centre. That was an important inflection point.
The first thing we decided was: we are going to make sure that customers always have a choice of using servers that are more efficient. Like in America, a lot of people drive big cars – they don’t need them, but they drive them. A lot of people like bigger power hungry servers over less power hungry ones. I think the first thing is to give customers the choice and to say ‘use this server, because here is how you can contribute to fight global warming. I think the basic unit of green is to engage the individual. So what we look to always do is to give the customers – there are some decisions that we make that is just the way of operating – others are about giving our customers the choice to do things for themselves. We are going to let them buy a smaller car instead of an SUV, so to speak. Giving them that choice lets the customer to say, ‘yes, I think I’m going to participate here.’
Sometimes, it is not appropriate for them, but we want to give customers the choice.
The second thing is we went and designed a data centre with very efficient air conditioning system. In a data centre, you have a lot of power being burnt in the servers, but every watt of that power goes to heat – electricity goes to either light or heat and there’s no light, so it all goes to heat. So if the server burns 300 watts, it’s all heat, so we have to take it out with air conditioning. It is a huge volume to remove. The way we remove it is one of the opportunities to be green. The way we do it is install air conditioning units that are heat exchangers. What that means is that you can either remove the heat through mechanical means through a compressor. Like the air conditioning in your house, you have a compressor that uses all the power. Ordinarily, you have a system that dissipates the heat and you have a compressor that compresses the refrigerant that discharges the heat. What we did was we brought a system that does this passively. So what a passive system allows you to do – say the temperature of the air inside is 22 degrees, if the air outdoor is 22 degrees, then all you have to do is take the air from outside and exchange it with the air from the inside.
This is not rocket science. This is not genius. It means that once we started paying attention, we can reap the benefits.
So is power a big cost component for Rackspace?
It’s a big expense, but it’s a small percentage. Power is less than 2% of our revenue. It’s still a large number. We are a company that last year, did US$360 million. Last quarter, we did US$130 million, so 2% of that is US$2.6 million, so roughly US$900,000 a month. It’s a big bill, but it is still a small percentage of our revenue because most of our revenue is for service. That is up substantially from last year – it’s a noticeable rise.
We want to be sure that we have an alternative for them and we can explain to them the clear advantage that brings them. It’s not really a cost thing for us, but a green thing. If we are managing a server for our customers for US$500, 2% of that is power, so US$10 – it’s not enough to matter. The green servers are usually more expensive.
There’s a lot of talk on power efficiency of data centres, what is your view on the topic?
Efficiency is work divided by energy. If you have a certain amount of processing power divided by the energy it uses, you could end up with an efficiency rating for the server.
On the data centre level, a lot of the power equals cooling, and redundancy. So we have a lot of systems that sit around and do nothing. If you want to make a data centre more efficiency, just make it less redundant. None of our customers are willing to take that trade off. The other problems is that if we put our data centres in Iceland it would be more efficient, it would be more efficient than if we put them in a hot place because cooling would be cheaper the colder you are. So if you are in Iceland, your ability to do passive cooling is much greater.
The problem is that it is hard to operate there. We are going to look at it, but I think it’s going to be a long shot. We have to be able to get servers there, generator experts there, but you never know, we’ll see.
Another issue is the power source. Is it coal? Is it hydro? Is it nuclear? That’s what makes it hard to have a standard. I would say, let’s have an Energy Star on routers, switches, UPSs, the air conditioning system itself – the things that draw energy. It’s like the air conditioner in your house. You can buy the really expensive one that is more efficient, or you can buy the cheap one that is not. When you make these decisions in the data centre, you end up with a more efficient data centre. The biggest factor is the efficiency of the server, which is hard to measure – in theory you can do it.
I think there’s a perception that data centres are a new problem. Say Rackspace manages 40,000 servers and that grows hundreds per month. So you can say, Rackspace, a million dollar of power, you are a polluter, etc. The dilemma of it is that we are taking the servers that the customers use to manage and managing it for them.
The largest users in the world still end up being companies who have a server room in the back in the office. What we do is transplant them from there to us. What we are doing is taking the servers that previously were in the server closet in the business and now we have them. So nothing has really happened. They had air conditioning in their offices that pulled the heat out – we have it.
What about cloud computing?
The server sits there most of the time, and spikes when there is an activity such as an email blast that you send out. Most of the time, it just sits there – this is the really wasteful thing.
The power demand rises obviously when the server is working harder, but only marginally. The ideal thing is to figure out a way to smooth out the demand, so we have the server running at full throttle all the time, because you can get more work done for the same amount of energy.
So the way you do that, is through cloud computing. It means it is pooled, it means it is shared – I don’t like the word share because it raises some questions and issues because of security, but pooled. It means that if you have 1000 servers and you have a Terabyte drive in each one, so you have 1000 Terabytes of data storage – that’s an unbelievable amount of storage and the whole city of Hong Kong probably didn’t have 1000 Terabytes of storage five years ago. Then you have dual processor with four cores, so you have eight cores per server – that’s 8000 cores.
These two resources can be used full throttle all of the time, or they can be used in a slight way. If they are used in the traditional way, the servers sit around and runs at 50%, 60%, 70% utilization with the air conditioning running, with the batteries being charged, with the routers running, everything has to run. It’s like leaving your car on idle all day long.
The real opportunity I think is, ‘look, instead of having a thousand servers doing this, doing a certain amount of work. How can we make the 1000 servers do more work?’ The answer is by using virtualization and cloud computing to level out the load. We think the load is around 4x. So the amount of work that can be done can be four times of what it is today using cloud computing vs the way it is done today.
We have two cloud computing services. One is email hosting. If you have a company that sends email for you, that server can be running full throttle all day. What we have, and I don’t have the exact number, is 400-500 servers that run 800,000 mail boxes. That means we are running 2,000 mail boxes per server. Think about the server in the average business, it is running 10 mail boxes.
So if we use mail boxes as a measure of work, say a hotel has 100 employees divided by one server, is productivity of 100. We have 2,000 mail boxes on a server, so that is 20x. Also, they are not running it all on one server, it’s a factory. Because if you want to run mail correctly, you need an inbound server, an outbound server, a virus server, a spam server and in some cases, there’s a Blackberry server, so it’s really five servers. So 100 divided by 5 is 20. So its between 20 per server, compared to 2,000, so its 100x.
What matters is utilization. You can theoretically come up with how many cycles a server produces and come up with a benchmark test, but what matters is utilization. So you want to be able to balance the load.
When we first started we had bandwidth that we had to pay for all day long, but that was only peaking at say, 7 pm, and the rest of the day, it was underutilised. So we opened up our UK office so we can sell the excess bandwidth during off-peak hours (because the peak hours for the UK is different from the peak hours in the US), because it’s free anyways. What we did was change the utilization curve to get more utilization out of it.
And this is the same basic idea with servers. Cloud computing, or pooled computing, will allow us to have the peak to trough utilization much better. Today, Rackspace has 40,000 servers, most of them dedicated to customers running in a curve. But if we can take the 40,000 servers and have them run like our bandwidth, we can do way more work. That’s the function of green, how we can use cloud computing.
Is this like virtualization?
Think about virtualization as the engine of a car. By itself it’s kind of useless, but if you put a car around it and put the power to the wheels then it works. The server itself has to run software on it. So in a normal software call the OS, then you run an application on top of that. In order to run Word on your PC, you need a processor, you need an OS to run on top of that, and you need Word to run on top of that. But if you have five people logged into a computer all running Word, it’s not going to work very well, because all five people are trying to shared the applications, and competing for resource – it’s like we are all trying to drink from the same cup of tea. You can do it, but I like to have my own cup.
Virtualization actually virtualises the server, so there is a layer that takes that server and splits it up into little server. These servers mean that we each get our own tea cup. We are going to have less, but we are all going to have our own.
So when you load Windows running Word, Windows is fooled into thinking that it has its own cup of tea, but what happens is when the other little servers are not running, the power gets transfer to the one that is running the applications. To Windows, it does not even know that it is running on a computer with other people. So it doesn’t even know that there’s a teapot and it is sharing the tea with other tea cups. All Windows knows is that it has its own tea cup.
What virtualization allows you to do is to take – it’s like cloud computer but on the server level – so instead of being one user on a server, you can have, say, 20 users. But they can’t all run at full throttle. The idea is to load-balance and get all of them to work and get utilization up.
So the first element of cloud computing is virtualization. The second part of it is, ‘here’s two servers in the data centre, and here’s all the little machines (the virtualized servers), and when one is filled up, they can moved it to the other server. At the end, the result is you get servers that run at 90%, instead of 20%. The thing is that all the software that developers have given us today is software that let you virtualize one server very well, but it won’t allow you to share between servers very well, and it won’t allow you to expand or contract this container.
My point really is that if you run in a cloud basis, it’s automatically more green. What we are doing is trying to take the capacity of the world today, how we can make the world’s servers more efficient.
The first important technology in cloud computing is the ability to create these virtual machines. But that is not really enough, because it only allows you to use multiple users on a single machine. The software that allows you to expand and contract the VM and the ability to move it across servers will be the next critical step. The key point is to have companies like Rackspace provide computing, so that more computing is provided by hosting companies – these will make computing more green.
One of the challenges is to convince companies to run applications on servers that are also running applications from other companies. But it’s not that hard a concept to get across. When you go to a bank, they don’t store your money in a corner by itself, they pool it, it’s the same concept.
So what’s next?
We have another offering called cloud FS, which is for file sharing. That’s in beta right now, and it is being launch. That is going to be a massive pool of storage. So if you think about your music, you upload your music to us and we’ll put it on some servers here, but the next time you back it up or upload, it will be on another group of servers, and so on.
The key is the software that will make it all look like one storage. That is the sort of thing that hosting companies are thinking about when delivering cloud computing services to customers.
We are the ones that have been delivering computing as a service for 10 years, and we are the ones developing the software to allow this to work.