|
Dynamic load balancing shall optimize the resource utilization in
parallel systems and computer networks in order to reduce the
application response times. While simple schemes are mathematically
tractable and can increase resource utilization for certain
applications, while implemented remote execution facilities can
avoid cpu overload peaks in computing clusters, load balancing
schemes for realistic systems and complex load profiles, especially
for database management systems, need to consider data and
communication.
This paper explains the problem and potential benefits of dynamic
load balancing, emphasizing on the exploitation of data affinities
and consideration of communication cost. The sophisticated
techniques developed in the HiCon system are presented in detail and
validated by several real measurements from parallel and distributed
database processing.
|