Use this URL to cite or link to this record in EThOS:
Title: Improving trade-offs between multiple metrics in parallel queueing systems
Author: Pesu, Tommi
ISNI:       0000 0004 7963 7900
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Parallel Queueing Networks can be used to model and optimise systems in many different en- vironments, such as distributed storage facilities, multi-core processors, RAID systems, supply chains and public services such as hospitals. The various stakeholders involved with the systems will often measure the performance of such systems using a wide range of metrics that often conflict with each other. Metrics of interest include task response time, subtask dispersion and energy consumption. Subtask dispersion is a recent metric, which is the difference in time of the first and last subtask to complete. The trade-offs between metrics can be controlled in various ways. Within this context, this thesis makes four primary contributions, the first of which of is to compare various delay- padding techniques in split-merge and fork-join parallel queueing models, with respect to task response time and subtask dispersion. We compare seven techniques from the literature, including some of our own, against each other across multiple case studies, in order to determine their strengths and weaknesses. Our results indicate that dynamic delay padding in a fork-join setting is currently the most promising technique for improving the trade-off between subtask dispersion and task response time. Our second contribution is to extend existing delay-padding techniques to work in a class of multi-layered parallel queueing environments, specifically Hidden Stochastic PERT Networks. We develop a technique which uses a state-of-the-art genetic algorithm to improve the trade- off between task service time and subtask dispersion. The method is able to robustly control subtask dispersion and task response time in a case study network. The third contribution is a systematic survey, which investigates alternative approaches for improving the performance of parallel queueing systems. Promising candidates include service restart and service replication. The final contribution is to combine multiple techniques: delay padding, state-dependent service strategies, service restart and service replication to provide greatly improved performance in terms of a three-way optimisation involving task response time, subtask dispersion and energy consumption. In the best case we managed to reduce the cost function by over 90% compared to an unoptimised system.
Supervisor: Knottenbelt, William Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral