Power utilization and heat dissipation are becoming increasingly important metrics for measuring the "health" of a data center. Current server racks can consume up to 18 kilowatts during periods of high server utilization, releasing much of this power as heat. Moreover, it is necessary to keep the server nodes within rated operating conditions. Approximately one-third of the power a data center consumes is used by air conditioners and fans, even when the majority of the server nodes are idle. For these reasons it is beneficial to correlate traditional system metrics with power and environmental conditions.
This talk will focus on an infrastructure at HP's Utility Data Center
that merges node utilization measurements with power and temperature
readings. I will discuss the components of the infrastructure, and how
we correlate the data. I will also present preliminary observations and
relations between performance and environmental data. I will conclude
with a roadmap for future areas of research, both at HP and as part of
the Cluster-on-Demand project at Duke.