A Framework for Improving Systems Performance by Minimizing Burstiness
Abstract
The principal goal of this work is to support performance management for systems that utilize resources in complex ways. Typically, performance evaluation has been carried out for such systems using simulation tools. However, such tools require expert model builders to create and maintain abstract business process models of the system under study. This can lead to a lack of representativeness, specifically, when many unique scenarios are to be modelled. This thesis presents a new simulation approach, Simulation By Example, which guides the simulation directly using events extracted from traces, i.e., examples. This work demonstrates and evaluates this new approach using three case studies from healthcare systems. These studies establish the advantages of SBE over traditional simulation methods and its ability to support a variety of performance management exercises.
Next, this thesis focuses on improving the performance of systems subjected to bursty workloads. Burstiness in resource service demands has previously been shown to have an adverse impact on system performance. This thesis proposes AMIR, an Analytic Method for Improving Responsiveness by reducing burstiness. AMIR considers a system with multiple classes of users and multiple resources that service user sessions in tandem. Batch processing systems, fabrication and manufacturing environments, micro-service systems, and patient operating rooms can be described in this way. Given the service demands distributions placed by all classes for the system's resources and the number of session arrivals for each class, AMIR decides an ordering of sessions that minimizes burstiness and improves system responsiveness metrics including session wait time, and total schedule processing time. A key aspect of the technique is an order O schedule burstiness metric β^O, which represents the mean joint probability that O+1 consecutive sessions in the schedule have resource demands at the bottleneck resource greater than the mean bottleneck resource demand of the schedule. For a given O, AMIR uses integer linear programming to produce schedules that progressively minimize β^i for all i in {1,...O}. Extensive simulation results show that AMIR significantly outperforms baseline policies such as shortest first and random scheduling. The results also provide insights on the conditions under which the technique is most effective.
Description
Keywords
Health Care Management, Computer Science, Engineering--System Science, Engineering--Operations Research
Citation
Kalbasi, A. (2017). A Framework for Improving Systems Performance by Minimizing Burstiness (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca. doi:10.11575/PRISM/24736