System Level Performance Evaluation of Distributed Embedded Systems
Khan, Subayal (2012)
Khan, Subayal
Tampere University of Technology
2012
Tieto- ja sähkötekniikan tiedekunta - Faculty of Computing and Electrical Engineering
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:ISBN:978-952-15-3241-2
https://urn.fi/URN:ISBN:978-952-15-3241-2
Tiivistelmä
In order to evaluate the feasibility of the distributed embedded systems in different application domains at an early phase, the System Level Performance Evaluation (SLPE) must provide reliable estimates of the nonfunctional properties of the system such as end-to-end delays and packet losses rate. The values of these non-functional properties depend not only on the application layer of the OSI model but also on the technologies residing at the MAC, transport and Physical layers. Therefore, the system level performance evaluation methodology must provide functionally accurate models of the protocols and technologies operating at these layers.
After conducting a state of the art survey, it was found that the existing approaches for SLPE are either specialized for a particular domain of systems or apply a particular model of computation (MOC) for modeling the communication and synchronization between the different components of a distributed application. Therefore, these approaches abstract the functionalities of the data-link, Transport and MAC layers by the highly abstract message passing methods employed by the different models of computation. On the other hand, network simulators such as OMNeT++, ns-2 and Opnet do not provide the models for platform components of devices such as processors and memories and totally abstract the application processing by delays obtained via traffic generators. Therefore the system designer is not able to determine the potential impact of an application in terms of utilization of the platform used by the device. Hence, for a system level performance evaluation approach to estimate both the platform utilization and the non-functional properties which are a consequence of the lower layers of OSI models (such as end-to-end delays), it must provide the tools for automatic workload extraction of application workload models at various levels of refinement and functionally correct models of lower layers of OSI model (Transport MAC and Physical layers).
Since ABSOLUT is not restricted to a particular domain and also does not depend on any MOC, therefore it was selected for the extension to a system level performance evaluation approach for distributed embedded systems. The models of data-link and Transport layer protocols and automatic workload generation of system calls was not available in ABSOLUT performance evaluation methodology. The, thesis describes the design and modelling of these OSI model layers and automatic workload generation tool for system calls. The tools and models integrated to ABSOLUT methodology were used in a number of case studies.
The accuracy of the protocols was compared to network simulators and real systems. The results were 88% accurate for user space code of the application layer and provide an improvement of over 50% as compared to manual models for external libraries and system calls. The ABSOLUT physical layer models were found to be 99.8% accurate when compared to analytical models. The MAC and transport layer models were found to be 70-80% accurate when compared with the same scenarios simulated by ns-2 and OMNeT++ simulators. The bit error rates, frame error probability and packet loss rates show close correlation with the analytical methods .i.e., over 99%, 92% and 80% respectively. Therefore the results of ABSOLUT framework for application layer outperform the results of performance evaluation approaches which employ virtual systems and at the same time provide as accurate estimates of the end-to-end delays and packet loss rate as network simulators. The results of the network simulators also vary in absolute values but they follow the same trend. Therefore, the extensions made to ABSOLUT allow the system designer to identify the potential bottlenecks in the system at different OSI model layers and evaluate the non-functional properties with a high level of accuracy. Also, if the system designer wants to focus entirely on the application layer, different models of computations can be easily instantiated on top of extended ABSOLUT framework to achieve higher simulation speeds as described in the thesis.
After conducting a state of the art survey, it was found that the existing approaches for SLPE are either specialized for a particular domain of systems or apply a particular model of computation (MOC) for modeling the communication and synchronization between the different components of a distributed application. Therefore, these approaches abstract the functionalities of the data-link, Transport and MAC layers by the highly abstract message passing methods employed by the different models of computation. On the other hand, network simulators such as OMNeT++, ns-2 and Opnet do not provide the models for platform components of devices such as processors and memories and totally abstract the application processing by delays obtained via traffic generators. Therefore the system designer is not able to determine the potential impact of an application in terms of utilization of the platform used by the device. Hence, for a system level performance evaluation approach to estimate both the platform utilization and the non-functional properties which are a consequence of the lower layers of OSI models (such as end-to-end delays), it must provide the tools for automatic workload extraction of application workload models at various levels of refinement and functionally correct models of lower layers of OSI model (Transport MAC and Physical layers).
Since ABSOLUT is not restricted to a particular domain and also does not depend on any MOC, therefore it was selected for the extension to a system level performance evaluation approach for distributed embedded systems. The models of data-link and Transport layer protocols and automatic workload generation of system calls was not available in ABSOLUT performance evaluation methodology. The, thesis describes the design and modelling of these OSI model layers and automatic workload generation tool for system calls. The tools and models integrated to ABSOLUT methodology were used in a number of case studies.
The accuracy of the protocols was compared to network simulators and real systems. The results were 88% accurate for user space code of the application layer and provide an improvement of over 50% as compared to manual models for external libraries and system calls. The ABSOLUT physical layer models were found to be 99.8% accurate when compared to analytical models. The MAC and transport layer models were found to be 70-80% accurate when compared with the same scenarios simulated by ns-2 and OMNeT++ simulators. The bit error rates, frame error probability and packet loss rates show close correlation with the analytical methods .i.e., over 99%, 92% and 80% respectively. Therefore the results of ABSOLUT framework for application layer outperform the results of performance evaluation approaches which employ virtual systems and at the same time provide as accurate estimates of the end-to-end delays and packet loss rate as network simulators. The results of the network simulators also vary in absolute values but they follow the same trend. Therefore, the extensions made to ABSOLUT allow the system designer to identify the potential bottlenecks in the system at different OSI model layers and evaluate the non-functional properties with a high level of accuracy. Also, if the system designer wants to focus entirely on the application layer, different models of computations can be easily instantiated on top of extended ABSOLUT framework to achieve higher simulation speeds as described in the thesis.
Kokoelmat
- Väitöskirjat [4862]