By now you've read a lot about green data centers. You've certainly heard all about IT's contribution to the growing need for power. But at the end of the day, what does it all mean? What is the reality of IT power consumption trends and how is that reality impacted by the core of the data center: the server?
I've spent most of my career asking these same questions. I drive systems-level energy strategy --- from the chip right on through the operating system layers and system software --- at Sun Microsystems. I'm helping ensure that our technologies and products are eco-responsible.
Earlier in my tenure with Sun, I was responsible for the technology direction for our blade servers, rack-mount servers, telecom-optimized servers, technical compute servers and workstations. I also drove the design and strategies for the next-generation system architectures, interconnects, and networking standards. Learning how to build efficient server architectures was great preparation for learning how to build efficient data centers.
So What's the Problem?
I think it's fair to say that we all understand the problem. The cost of power as a percentage of overall IT spending ranges from 20 percent for a small enterprise data center to about 80 percent for very large hosting and collocation providers. And if you deconstruct data center power needs further, it becomes clear that the largest resource hogs are the servers, storage and switch infrastructure---ironically, equaled by the resources required of the cooling units to keep the temperatures of these rack-mount devices in check.
Given the rising cost of electricity, it is estimated that the power and cooling costs required to run a low-cost server may outweigh the acquisition cost of that server in the very near future. That's not great news. However, I believe 2008 is a watershed year for energy-efficiency in IT. The environmental impacts of massive, global data centers make eco-responsibility more important than ever.
Light at the End of the Hot Aisle
I don't know about your environment, but many of our customers see no end to the pressures to deploy more applications and support greater numbers of users. New applications and the data resources to support them are squeezed into the existing data center footprint, compounding the problem. We're seeing a space, power and cooling crunch born out of the need for more, bigger and faster resources.
This modus operandi is further complicated by the silo approach to server resource sizing, where application server needs are sized to handle the maximum need at any given time. Extra server capacity simply sits idle, consuming resources.
But there is an end in sight. As an engineer, I'm thrilled at the progress we're making as an industry in a short amount of time.
The challenge is to build energy-efficient servers that provide superior computing power but require less energy to run --- and to manufacture, too. That goal requires a relentless focus on delivering performance per watt at every level of the stack. In the past, our industry focused on producing faster and higher-performing servers, with little regard for their power consumption. The tide is starting to change industry-wide, though, as more vendors emphasize increased energy efficiency, without sacrificing either performance or reliability.
At the Core -- Enter the Energy-Efficient Server
I believe that the solution to building more efficient data centers is two-fold: Use more powerful, eco-responsible servers, while at the same time, increase the overall efficiency of the data center by improving power distribution and cooling.
As a rule of thumb, for every watt delivered to power today's IT equipment, at least one additional watt is spent on overhead, including power and cooling resources. We can have a dramatic impact on overall power consumption by reducing the load and increasing the efficiency of power and cooling.
In the server world, getting more power efficiencies from your system resources requires a focused effort in the areas of processor, memory, I/O subsystem, power and cooling. If you start at the core with good green design in mind, your chips will consume less power per core and thread. The CPU is one of the most power hungry components of today's server; advances here will pay dividends when amplified throughout the rack and across the data center floor. A watt saved in server power will yield an additional watt (or two) of savings in cooling power as well.
For example, by leveraging a simple CPU core architecture, we can replicate the core many times over on a single piece of silicon. Then, by adding a number of vertical threads on each core, we can increase its utilization even further. Consider the recent evolution that began with the first 8-core general-purpose CPU with 4 threads each (giving it the power of 32 microprocessors) on a single die. Later we saw CPUs with 8 cores and 8 threads per core, and in 2008, we are seeing scalable variants of these systems with multiple CPUs per box.
The beauty of this design is that because each tick of the CPU clock causes many things to happen to parallel on all the available cores, the CPU doesn't have to be clocked very fast. And not clocking it fast means it doesn't consume a lot of power. Like another architectural innovation --- the hybrid car --- the CPU gives you a lot more mileage while sipping only a tiny amount of fuel. Using this modular approach to design not only provides for greater performance and compatibility, but also saves on power.
But there's also a softer side to this problem. Software technologies, such as virtualization, provide for the ability to take a single environment and create the illusion of multiple environments. This ability has been around for a number of years and can be implemented at many different levels within the overall system and data center.
For example, virtualization technologies can provide for virtual servers within a single server enclosure; where each has its own dedicated processor, memory and I/O resources. Now we can begin to break down the silos and consolidate multiple workloads into a single server, resulting in more efficient use of each server resource while at the same time creating a reduction in the overall power demand. Virtualization is now fundamental to maximizing data center efficiency.
We'll leave disk and I/O subsystem for another discussion, but let's not overlook other areas of system architecture, namely memory, power and cooling. As processor speeds and the need for memory-hungry applications (think virtualization) have increased, so has the need for larger memory subsystems and faster DIMMs. Again, designing with this in mind will yield greater efficiencies down the road. For example, building servers with larger numbers of DIMM sockets per processor will allow technologies such as virtualization to have a greater effect on server consolidation. It's obvious that bigger servers can accommodate more virtual machines per physical server --- thereby leading to greater overall utilization and much greater efficiency per watt. It's the same logic by which a fully occupied passenger bus yields a much better ratio of gallons per passenger mile than a fully occupied sedan (which can go faster) or a single-occupancy Formula 1 car (which can go even faster).
What Does the Future Hold?
The future of energy-efficient data centers and high performance computing environments is one that must be approached holistically. It involves a plan that is centered around a new approach to system design and one that embraces the ecosystem.
One component of this plan will involve a new design approach that includes dynamic power and control --- in short, pushing more intelligence down to the hardware system level. Think about optimized power and cooling; a server architecture that will automatically slow the fan and CPU clock speed when certain environmental triggers are present. I envision a level of server intelligence that will self-analyze and tune its power and cooling characteristics to suit the immediate need, independent of the OS.
Another aspect of this plan will center around working more closely with partners along the power chain. I advocate that server manufacturers, cooling technology manufacturers, system integrators, consultants and more need to work closely to ensure that we enter this next phase of the data center revolution prepared.
This plan also will involve building servers that respond to the specific energy policy that the customer wants to impose on them. If the customer wants to limit the server to a particular power budget during a utility brownout, or the customer wants to calibrate the server's power draw to be proportional to the predictable fluctuations in its daily workload, they should be able to do so. Just like you might choose to drive a car in low gear going uphill to manage engine load, in high gear on the freeway for speed, and in neutral coming downhill to save fuel, you should be able to drive a server in whatever gear you want that is appropriate to your computing workload.
In closing, we're now at the point where we understand that maximizing energy efficiency is not just a chip or hardware problem. It's a virtualization problem, an OS problem, a systems management problem, a network problem and a storage problem. The solution will require a more proactive, holistic approach to building systems.
Subodh Bapat is Vice President and Distinguished Engineer in the Eco-Responsibility Office at Sun Microsystems.