Dominant Resource Fairness: Fair Allocation of Heterogeneous Resources in Datacenters
Ali Ghodsi, Matei Zaharia, Benjamin Hindman, Andrew Konwinski, Scott Shenker and Ion Stoica
EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2010-55
May 7, 2010
http://www.eecs.berkeley.edu/Pubs/TechRpts/2010/EECS-2010-55.pdf
This paper investigates how different resources can be fairly allocated among users that possibly prioritize them differently. We introduce a fairness policy, called Dominant Resource Fairness (DRF), which is an adaptation of max-min fairness from networking to datacenter environments. We show that DRF, unlike other policies which we investigated, satisfies a number of desirable properties that a fair datacenter scheduler should have, including guaranteeing that every user gets 1/N of some resource and that users can relinquish resources without hurting other users’ allocations. DRF is also envy-free, incentivizing users to correctly report their resource demand. When compared to other intuitive schedulers, as well as competing ones from microeconomic theory, DRF is more fair.
BibTeX citation:
@techreport{Ghodsi:EECS-2010-55,
Author = {Ghodsi, Ali and Zaharia, Matei and Hindman, Benjamin and Konwinski, Andrew and Shenker, Scott and Stoica, Ion},
Title = {Dominant Resource Fairness: Fair Allocation of Heterogeneous Resources in Datacenters},
Institution = {EECS Department, University of California, Berkeley},
Year = {2010},
Month = {May},
URL = {http://www.eecs.berkeley.edu/Pubs/TechRpts/2010/EECS-2010-55.html},
Number = {UCB/EECS-2010-55},
Abstract = {This paper investigates how different resources can be fairly allocated among users that possibly prioritize them differently. We introduce a fairness policy, called Dominant Resource Fairness (DRF), which is an adaptation of max-min fairness from networking to datacenter environments. We show that DRF, unlike other policies which we investigated, satisfies a number of desirable properties that a fair datacenter scheduler should have, including guaranteeing that every user gets 1/N of some resource and that users can relinquish resources without hurting other users’ allocations. DRF is also envy-free, incentivizing users to correctly report their resource demand. When compared to other intuitive schedulers, as well as competing ones from microeconomic theory, DRF is more fair.}
}
EndNote citation:
%0 Report %A Ghodsi, Ali %A Zaharia, Matei %A Hindman, Benjamin %A Konwinski, Andrew %A Shenker, Scott %A Stoica, Ion %T Dominant Resource Fairness: Fair Allocation of Heterogeneous Resources in Datacenters %I EECS Department, University of California, Berkeley %D 2010 %8 May 7 %@ UCB/EECS-2010-55 %U http://www.eecs.berkeley.edu/Pubs/TechRpts/2010/EECS-2010-55.html %F Ghodsi:EECS-2010-55