You say you want a revolution? Energy efficient data centers are in the news again, with the EPA reporting that data centers use 1.5% of US electricity - almost 6 million home's worth - and doubling in five years.
But what about your home data center? My house has 2 wireless routers, 3 systems, 4 monitors, 6 cores and 8 GB RAM. They don't use as much power as a refrigerator, but I'd like them to be more efficient anyway.
Measuring & comparing power use Can we cut the power requirement? We could, if we had a way to reliably benchmark power consumption across architectures. Which is what JouleSort: A Balanced Energy-Efficiency Benchmark (PDF) by Suzanne Rivoire, Mehul A. Shah, Parthasarathy Ranganathan and Christos Kozyrakis tries to do.
The benchmark of the future The authors chose a sort algorithm that would exercise the entire system:
JouleSort is an I/O-centric benchmark that measures the energy efficiency of systems at peak use. Like previous sort benchmarks, one of its goals is to gauge the end-to-end effectiveness of improvements in system components. To do so, JouleSort allows us to compare the energy efficiencies of a variety of disparate system configurations. . . . previous sort benchmarks have been technology trend bellwethers, for example, foreshadowing the transition from supercomputers to clusters. Similarly, an important purpose of JouleSort is to chart past trends and gain insight into future trends in energy efficiency.
Prototyping an energy efficient server They used the benchmark to evaluate several systems, some "unbalanced" systems such as a laptop and systems "balanced" or configured to meet the needs of the benchmark most efficiently.
They found that unbalanced CPU utilization was quite low, ranging from 1% to 26%. As a result, the system didn't accomplish much work for the power it consumed.
Since the CPU is usually the highest power component, these results suggest that building a system with more I/O to complement the available processing capacity should provide better energy efficiencies.
Ah, the irony! 40 years after the minicomputer we are back to a batch mainframe I/O-centric architecture. All things old are new again.
Design for efficiency Disks and bandwidth are critical for efficiency in this benchmark. At 15 W each, it doesn't take many SATA disks to overtake the CPU as the major power sink. The balanced system required 2 trays of 6 disks each to keep a dual-core CPU busy.
A really efficient server The team then built a server optimized for the benchmark. As I noted in Power pushing 2.5? drives to tipping point 2.5" notebook drives are 2-3 watts, not 10-15. And sure enough, the paper found that an energy efficient system did much better with notebook drives. The configuration:
File systems, RAM and power supplies The benchmark is a sequential sort. The authors found that file systems with higher sequential access rates more efficient. Developers, the time may not too far away when your code is measured on power efficiency as well as performance.
They also found that reducing the RAM footprint to the needed capacity raised efficiency as well. FB-DIMMs, with their 5 watt per stick penalty, are a definite efficiency target.
They also found that the winning system could use a much smaller power supply. The authors suggest that power-factor corrected power supplies are required to make energy efficient servers economic as well.
The StorageMojo take As the breadth of the paper suggests, power efficiency requires a holistic understanding of computes, I/O, software, power factors and configuration trade-offs. Some of the supercomputer folks can do this, but the average data center or home user is years away from this level of workload understanding.
Instead the research points to a few things that increase efficiency and reduce consumption across a wide range of workloads and configurations. Mobile CPUs and notebook disks are 2 likely candidates. Software effects are significant as well because widely used software affects so many systems.
Will we all be buying home servers with a half dozen disks any time soon? No. But saving power while increasing performance is always a good thing.
Comments welcome, as always. For home users power cost is not a big issue. How much more would you be willing to pay for a more efficient system?