Laura Keys

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2010-117

August 18, 2010

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2010/EECS-2010-117.pdf

Traditional servers account for more than 1.5\% of the US electricity use though spend their lives largely underutilized or idle. Because a large portion of power in a data center is due directly or indirectly to servers, power savings in a data center environment can be achieved simply by using lower power hardware in place of these traditional servers. However, deciding which hardware to use in place of servers is complicated because lower power typically equates to lower performance and because different cluster owners use different metrics of success in quantifying cluster performance. In this project report I present measurements from several single-machine and system benchmarks for both interactive and batch jobs and develop predictive models for power and performance within a cluster. Accurately predicting power to within 10\% based on OS-reported metrics requires similar training and testing benchmarks, while predicting performance within a heterogeneous web server is simple and straightforward. I also introduce a Cluster Visualizer for comparing potential cluster configurations based on the actual benchmark measurements and different metrics of value, ultimately making the decision about which hardware platform to build a cluster from less complicated.

---

BibTeX citation:

@techreport{Keys:EECS-2010-117,
    Author= {Keys, Laura},
    Title= {A Model-Based Process for Evaluating Cluster Building Blocks},
    Year= {2010},
    Month= {Aug},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2010/EECS-2010-117.html},
    Number= {UCB/EECS-2010-117},
    Abstract= {Traditional servers account for more than 1.5\% of the US electricity use though spend their lives largely underutilized or idle. Because a large portion of power in a data center is due directly or indirectly to servers, power savings in a data center environment can be achieved simply by using lower power hardware in place of these traditional servers. However, deciding which hardware to use in place of servers is complicated because lower power typically equates to lower performance and because different cluster owners use different metrics of success in quantifying cluster performance. In this project report I present measurements from several single-machine and system benchmarks for both interactive and batch jobs and develop predictive models for power and performance within a cluster. Accurately predicting power to within 10\% based on OS-reported metrics requires similar training and testing benchmarks, while predicting performance within a heterogeneous web server is simple and straightforward. I also introduce a Cluster Visualizer for comparing potential cluster configurations based on the actual benchmark measurements and different metrics of value, ultimately making the decision about which hardware platform to build a cluster from less complicated.},
}

EndNote citation:

%0 Report
%A Keys, Laura 
%T A Model-Based Process for Evaluating Cluster Building Blocks
%I EECS Department, University of California, Berkeley
%D 2010
%8 August 18
%@ UCB/EECS-2010-117
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2010/EECS-2010-117.html
%F Keys:EECS-2010-117