The challenges in building energy efficient clusters of
servers with low-power CPUs are two fold: (1) balancing I/O resource utilization with power consumption, and
(2) amortizing the power consumption of fixed server components like fans and power supplies. In this work we
show that the I/O resource utilization and power consumption can be balanced by using a virtualized I/O approach.
The fixed component power overheads on the other hand
can be amortized by sharing them among many servers
with low-power CPUs. We demonstrate a system built on
these ideas and study its characteristics for web-class applications. The system shows significant improvements in
Perf./W and Perf./W-h for Hadoop applications and scales
to a large number of cluster nodes. Compared to a commodity cluster using heavyweight CPUs in the same power
budget, our system delivers 3X average improvement in
Perf./W-h, and 64% average improvement in Perf./W. For
a fixed power budget, the execution time also improves by
57% compared to a traditional cluster for benchmarks that
showed improved performance.