Performance Studies of Dynamic Load Balancing in Distributed Systems
Abstract
Distributed systems are often characterized by uneven loads on hosts and other resources. In this thesis, the problems concerning dynamic load balancing in loosely-coupled distributed systems are studied using trace-driven simulation, implementation, and measurement. Information about job CPU and input/output demands is collected from three production systems and used as input to a simulator that includes a representative central processing units scheduling policy and considers the message exchange and job transfer costs explicitly. A prototype load balancer is implemented in the Berkeley UNIX and Sun/UNIX environments, and the results of a large number of measurement experiments performed on six workstations are presented. The quality of two families of load indices, one based on resource queue length, the other on resource utilization, is evaluated in the context of dynamic load balancing. The performances of seven algorithms using different load information exchange and job placement strategies are compared. The factors that affect load balancing performance, and the impacts of load balancing on individuals hosts and on each type of job are also quantitatively investigated.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 1987
- Accession Number
- ADA197130
Entities
People
- Songnian Zhou
Organizations
- University of California, Berkeley