[Back] [PDF]

We Didnít Start the Fire: Using Agent-Directed Thermal Modeler to Keep Servers Cool


J. Hood, T. Scott, J. Yu, X. Qin*, L. Yilmaz, and J. A. Hamilton


Computer Science and Software Engineering

Auburn University, Auburn, AL 36849

{ fjwh0011, tas0015, jzy0012, xqin, yilmale}@eng.auburn.edu


As energy use by datacenters has risen over the years, the costs required to run a datacenter have substantially increased. Several algorithms for thermal management and thermal-aware job placement exist; however, choosing the scheme that will most efficiently cool a datacenter can be challenging. Thermal models offer a great solution to help choose which algorithm will perform best by juxtaposing different thermal-aware algorithms. When the temperature can be observed over all servers through simulated steps, one can decipher the differences and advantages of one thermal-aware algorithm over another. Existing thermal modelers, however, can be slow and may take a while to learn to use. When one wishes to compare several thermal models, waiting for hours for the result of one thermal-aware algorithm may mean that not as many algorithms can be compared. Agent-Directed Thermal Modeler (ADTM) provides a solution that has a low learning curve and still produces visualizations of datacenters quickly. It is aimed at being easy to configure while still producing very meaningful data through graphs and images. There are very few parameters, making setup much easier - one simply has to configure a few settings such as initial number of jobs, job gain, and which thermal-aware algorithm to use. Once these settings are configured, the simulation can be run in a matter of seconds. A single time-step of the simulation takes milliseconds. The graphical and pictorial output of ADTM can then be used to determine which thermal-aware algorithm works best for a given datacenter in a much shorter time than other thermal-modelers. ADTM is used to compare XInt-GA to random job placement in order to show that a 5% increase in energy savings is expected in an overloaded datacenter. This simulation for both algorithms takes only a few seconds, so many thermal-aware algorithms can be compared quickly in order to determine the most effective and realistic algorithm that can be chosen to cool a datacenter.

This paper appeared in the Proceedings of the Agent-Directed Simulation Symposium of the SpringSim Multiconference, Tampa, Florida, April 2014.

*Corresponding author. http://www.eng.auburn.edu/~xqin


Acknowledgments: This research was supported by the U.S. National Science Foundation under Grants CCF-0845257 (CAREER), CNS-0917137 (CSR), CNS-0757778 (CSR), CCF-0742187 (CPA), CNS-0831502 (CyberTrust), CNS-0855251 (CRI), OCI-0753305 (CI-TEAM), DUE-0837341 (CCLI), and DUE-0830831 (SFS).