Domenico Ferrari and Songnian Zhou

EECS Department
University of California, Berkeley
Technical Report No. UCB/CSD-87-353
May 1987

http://www2.eecs.berkeley.edu/Pubs/TechRpts/1987/CSD-87-353.pdf

In this paper, we empirically evaluate the quality of several load indices in the context of dynamic load balancing. We have implemented a load balancer for Sun/UNIX environments. In our experimental setup, six Sun-2 workstations were driven by job scripts, and job response times were measured while loads were being balanced and various load indices used to make job placement decisions. We study the effects on performance of the choice of load index, the averaging interval, the load information exchange period, and the characteristics of the workload. Measurements show that the performance benefits of load balancing are indeed strongly dependent upon the load index. Load indices based on resource queue lengths are found to perform better than those based on resource utilization, and the use of an exponential smoothing method yields further improvement over that of instantaneous queue lengths.

BibTeX citation:

@techreport{Ferrari:CSD-87-353,
    Author = {Ferrari, Domenico and Zhou, Songnian},
    Title = {An Empirical Investigation of Load Indices for Load Balancing Applications},
    Institution = {EECS Department, University of California, Berkeley},
    Year = {1987},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/1987/5990.html},
    Number = {UCB/CSD-87-353},
    Abstract = {In this paper, we empirically evaluate the quality of several load indices in the context of dynamic load balancing. We have implemented a load balancer for Sun/UNIX environments. In our experimental setup, six Sun-2 workstations were driven by job scripts, and job response times were measured while loads were being balanced and various load indices used to make job placement decisions. We study the effects on performance of the choice of load index, the averaging interval, the load information exchange period, and the characteristics of the workload. Measurements show that the performance benefits of load balancing are indeed strongly dependent upon the load index. Load indices based on resource queue lengths are found to perform better than those based on resource utilization, and the use of an exponential smoothing method yields further improvement over that of instantaneous queue lengths.}
}

EndNote citation:

%0 Report
%A Ferrari, Domenico
%A Zhou, Songnian
%T An Empirical Investigation of Load Indices for Load Balancing Applications
%I EECS Department, University of California, Berkeley
%D 1987
%@ UCB/CSD-87-353
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/1987/5990.html
%F Ferrari:CSD-87-353