Parametric Estimation of Load for Air Force Data Centers
Abstract
The Office of Management and Budget (OMB) has tasked Federal agencies to develop a Data Center Consolidation Plan. Effective planning requires a repeatable method to effectively and efficiently size Air Force Base-level data centers. Review of commercial literature on data center design found emphasis in power efficiency, thermal modeling and cooling, and network speed and availability. The topic of sizing data center processing capacity seems undeveloped. This thesis provides a better, pedigreed solution to the data center sizing problem. By analogy, Erlang's formulae for the probability of blocking and queuing should be applicable to cumulative CPU utilization in a data center. Using survey data collected by 38th Engineering Squadron, a simulation is built and correlation between the observed survey measurements and simulation measurements, and the Erlang, Gamma, and Gaussian-Normal distributions is found. For a sample dataset of 70 servers over 14 hours of observation and a supposed .99999 requirement for traffic to be passed or otherwise unimpeded, Erlang distribution predicts 10 CPU cores are required, Gamma distribution predicts 10 CPU cores are required, Gaussian-Normal distribution predicts 9 CPU cores are required, Erlang B formulae predicts 14 CPU cores are required, and Erlang C formulae predicts 15 CPU cores are required.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 27, 2015
- Accession Number
- ADA616254
Entities
People
- Derek P. Molle
Organizations
- Air Force Institute of Technology