Dean Nelson, the senior director at eBay Inc., believes in the power of the data center operator to change the industry for the better. In 2008, the longtime Sun Microsystems data center manager co-founded Data Center Pulse (DCP), a grassroots end user-only organization., because he was frustrated by traditional user groups' inability to push vendors to change and improve their products.
In 2009, DCP published its
top 10 demands of the data center industry and the results of independent vendor bakeoffs for more efficient cooling strategies. While these projects have moved the industry incrementally forward, Nelson and DCP's next project could turn data center design on its head.
How much influence has Data Center Pulse had on your career?
Dean Nelson: I spend a lot of time at work --we all do. And I want to spend that time wisely.
[Data center managers] were not getting what we needed out of the industry. Companies were coming out with me-too product solutions and the end-user community was duplicating effort all over the place. So by bringing users together, that voice is starting to come into one. I feel good about that. I'm benefitting from a peer network and great friendships. But we have the influence to do things. The IT economy is $336 billion. We consume a lot of that -- money coming out of our individual companies -- whether it's buying hardware or paying the power bill. And that voice of customers can have a lot of influence.
Have you seen companies change direction or product lines due to Data Center Pulse?
D.N.: Vendors are using our top 10 list. What I love is [that] people use it in presentations and to justify product development. That has influenced the industry. We have also been pushing the industry with our Chill Off projects.
In the Chill Off, you test data center cooling approaches to see which is most efficient --
right? You're now onto the third round: Chill Off Three. How does it work?
D.N.: I want to answer one question: Is water cooling going to give me a performance boost in my IT capacity?
We're all using air to cool our servers, and it's terribly inefficient. So how do we go back and do something different? The chips are optimized for air cooling; they're not optimized for the best performance. Server vendors literally hold the frequency of the CPU to a certain level because it just gets too hot.
For certain applications like search, the workload fits in server memory. So the frequency of the CPU is directly proportional to performance. If CPU frequency goes up, I need less boxes to do the search. So if I can use liquid cooling and increase clock frequency, I can cut my box count. I can get more performance out of my system.
We're going to use "queries per watt" as a metric. We're using eBay's search engine architecture for the test. I don't care what the cooling design is or the hardware that goes into it. It comes down to how many queries will you give me for every watt I consume? That means facilities and IT have to work together on one type of system to give me that output.
At eBay we've raised the power distribution voltage, built cooling containment, and implemented all other best practices. Every two years, we're replacing our servers so we're already cutting our power consumption in half and doubling performance. But how do we get more? It may be by overclocking CPUs to get that major jump. I want to see if there's a quantum leap.
What's more fun: Building out a data center or tearing one down?
Nelson: Decommissioning is hard. Consolidation is really tough. You're trying to take stuff from a flying airplane and moving it to another flying airplane. You're rebuilding an engine while it's running. The shutdown of the data center on the EBay site was a really significant thing. They'd never shut one down. They've always built more. That was the trend, and now we've broke that. We just can't sustain that.
Building is always fun. You can do everything new, always the right way. And if you own the actual budget spreadsheet and architecture, you can do amazing things. When you don't, that's when the battles and challenges come in and it becomes really painful to build something out. You need to leverage and partner with people to spend money on things. I probably like building out better than tearing down.
Some have suggested that large Internet companies can move from the physical redundancy of
data centers and save power. Microsoft's Christian Belady has said that some of the company's new
data centers don't even have uninterruptible power supply systems. What are the possibilities of
doing failover with software technology rather redundant data center components?
D.N.: I think it can drive efficiency, but it won't be 100%. There will always be core workloads you can't fail over. If we have a fault, we can fail over complete data centers, but we will take a hit because state information must transfer. We're not active -- so that within a second, I have my 1.3 million people's transactions captured and instantaneously moved.
Christian is right. Search is search. It's replicated. If this one fails, I just point over there. But [the] back end of search -- [the] databases and the other components that enable it to happen? That's a whole other story. I absolutely support virtualization happening everywhere, virtualization will allow us that agility to move workloads between data centers. All of that is happening, and I think that will mature over time and we'll see some significant abilities for certain workloads to fail over at any time. That's the whole cloud concept. But that will never be 100%. That's my take today, but next year we may have the next innovation and the possibility of something new.
Let us know what you think about the story; email Matt Stansberry, Executive Editor.