|
Title
|
|
|
|
Queueing with flexible and heterogeneous servers
| |
|
Author
|
|
|
|
| |
|
Abstract
|
|
|
|
| Load balancing is one of the key components in many distributed systems as it heavily impacts performance and resource utilization. This thesis considers heterogeneous systems where servers belong to one of multiple classes, with the speed of an individual server depending on its class. Two categories of load balancing strategies are considered. The first category consists of centralized load balancing strategies, where a dispatcher assigns incoming jobs to the servers. Both randomized dispatching and size-based dispatching strategies are considered. For these strategies, we propose mathematical models to describe the queue lengths over time and validate them through discrete event simulation. We show that the dispatcher's optimal parameter can be determined via convex optimization, if it exists. Furthermore we also investigate the effects of different system parameters on the achievable mean response time and propose several easier to compute schemes for determining the dispatcher's selection parameter. Finally, we also discuss the impact of scheduling policies in the nodes in addition to the load balancing strategies. The second category consists of decentralized load balancing strategies where no dispatcher is used, but servers can exchange jobs through communication by means of probing. For these load balancing strategies, we use a mean field model to study the queue lengths and the required probe rate to achieve stability. We also develop an iterative algorithm to more easily compute the mean response time. Finally we study and compare the mean response times for specific "pull" and "push" strategies for different parameter settings. |
| |
|
Language
|
|
|
|
English
| |
|
Publication
|
|
|
|
Antwerp
:
University of Antwerp
,
2020
| |
|
Volume/pages
|
|
|
|
127 p.
| |
|
Note
|
|
|
|
:
Van Houdt, B. [Supervisor]
| |
|
Full text (open access)
|
|
|
|
| |
|