Observation mechanisms for in-field software-based self-test
Supervisor(es): Sonza Reorda, Matteo - Canetti, Rafael
Resumen:
When electronic systems are used in safety critical applications, as in the space, avionic, automotive or biomedical areas, it is required to maintain a very low probability of failures due to faults of any kind. Standards and regulations play a significant role, forcing companies to devise and adopt solutions able to achieve predefined targets in terms of dependability. Different techniques can be used to reduce fault occurrence or to minimize the probability that those faults produce critical failures (e.g., by introducing redundancy). Unfortunately, most of these techniques have a severe impact on the cost of the resulting product and, in some cases, the probability of failures is too large anyway. Hence, a solution commonly used in several scenarios lies on periodically performing a test able to detect the occurrence of any fault before it produces a failure (in-field test). This solution is normally based on forcing the processor inside the Device Under Test to execute a properly written test program, which is able to activate possible faults and to make their effects visible in some observable locations. This approach is also called Software-Based Self-Test, or SBST. If compared with testing in an end of manufacturing scenario, in-field testing has strong limitations in terms of access to the system inputs and outputs because Design for Testability structures and testing equipment are usually not available. As a consequence there are reduced possibilities to activate the faults and to observe their effects. This reduced observability particularly affects the ability to detect performance faults, i.e. faults that modify the timing but not the final value of computations. This kind of faults are hard to detect by only observing the final content of predefined memory locations, that is the usual test result observation method used in-field. Initially, the present work was focused on fault tolerance techniques against transient faults induced by ionizing radiation, the so called Single Event Upsets (SEUs). The main contribution of this early stage of the thesis lies in the experimental validation of the feasibility of achieving a safe system by using an architecture that combines task-level redundancy with already available IP cores, thus minimizing the development time. Task execution is replicated and Memory Protection is used to guarantee that any SEU may affect one and only one of the replicas. A proof of concept implementation was developed and validated using fault injection. Results outline the effectiveness of the architecture, and the overhead analysis shows that the proposed architecture is effective in reducing the resource occupation with respect to N-modular redundancy, at an affordable cost in terms of application execution time. The main part of the thesis is focused on in-field software-based self-test of permanent faults. A set of observation methods exploiting existing or ad-hoc hardware is proposed, aimed at obtaining a better coverage, in particular of performance faults. An extensive quantitative evaluation of the proposed methods is presented, including a comparison with the observation methods traditionally used in end of manufacturing and in-field testing. Results show that the proposed methods are a good complement to the traditionally used final memory content observation. Moreover, they show that an adequate combination of these complementary methods allows for achieving nearly the same fault coverage achieved when continuously observing all the processor outputs, which is an observation method commonly used for production test but usually not available in-field. A very interesting by-product of what is described above is a detailed description of how to compute the fault coverage achieved by functional in-field tests using a conventional fault simulator, a tool that is usually applied in an end of manufacturing testing scenario. Finally, another relevant result in the testing area is a method to detect permanent faults inside the cache coherence logic integrated in each cache controller of a multi-core system, based on the concurrent execution of a test program by the different cores in a coordinated manner. By construction, the method achieves full fault coverage of the static faults in the addressed logic.
Cuando se utilizan sistemas electrónicos en aplicaciones críticas como en las áreas biomédica, aeroespacial o automotriz, se requiere mantener una muy baja probabilidad de malfuncionamientos debidos a cualquier tipo de fallas. Los estándares y normas juegan un papel importante, forzando a los desarrolladores a diseñar y adoptar soluciones que sean capaces de alcanzar objetivos predefinidos en cuanto a seguridad y confiabilidad. Pueden utilizarse diferentes técnicas para reducir la ocurrencia de fallas o para minimizar la probabilidad de que esas fallas produzcan mal funcionamientos críticos, por ejemplo a través de la incorporación de redundancia. Lamentablemente, muchas de esas técnicas afectan en gran medida el costo de los productos y, en algunos casos, la probabilidad de malfuncionamiento sigue siendo demasiado alta. En consecuencia, una solución usada a menudo en varios escenarios consiste en realizar periódicamente un test que sea capaz de detectar la ocurrencia de una falla antes de que esta produzca un mal funcionamiento (test en campo). En general, esta solución se basa en forzar a un procesador existente dentro del dispositivo bajo prueba a ejecutar un programa de test que sea capaz de activar las posibles fallas y de hacer que sus efectos sean visibles en puntos observables. A esta metodología también se la llama auto-test basado en software, o en inglés Software-Based Self-Test (SBST). Si se lo compara con un escenario de test de fin de fabricación, el test en campo tiene fuertes limitaciones en términos de posibilidad de acceso a las entradas y salidas del sistema, porque usualmente no se dispone de equipamiento de test ni de la infraestructura de Design for Testability. En consecuencia se tiene menos posibilidades de activar las fallas y de observar sus efectos. Esta observabilidad reducida afecta particularmente la habilidad para detectar fallas de performance, es decir fallas que modifican la temporización pero no el resultado final de los cálculos. Este tipo de fallas es difícil de detectar por la sola observación del contenido final de lugares de memoria, que es el método usual que se utiliza para observar los resultados de un test en campo. Inicialmente, el presente trabajo estuvo enfocado en técnicas para tolerar fallas transitorias inducidas por radiación ionizante, llamadas en inglés Single Event Upsets (SEUs). La principal contribución de esa etapa inicial de la tesis reside en la validación experimental de la viabilidad de obtener un sistema seguro, utilizando una arquitectura que combina redundancia a nivel de tareas con el uso de módulos hardware (IP cores) ya disponibles, que minimiza en consecuencia el tiempo de desarrollo. Se replica la ejecución de las tareas y se utiliza protección de memoria para garantizar que un SEU pueda afectar a lo sumo a una sola de las réplicas. Se desarrolló una implementación para prueba de concepto que fue validada mediante inyección de fallas. Los resultados muestran la efectividad de la arquitectura, y el análisis de los recursos utilizados muestra que la arquitectura propuesta es efectiva en reducir la ocupación con respecto a la redundancia modular con N réplicas, a un costo accesible en términos de tiempo de ejecución. La parte principal de esta tesis se enfoca en el área de auto-test en campo basado en software para la detección de fallas permanentes. Se propone un conjunto de métodos de observación utilizando hardware existente o ad-hoc, con el fin de obtener una mejor cobertura, en particular de las fallas de performance. Se presenta una extensa evaluación cuantitativa de los métodos propuestos, que incluye una comparación con los métodos tradicionalmente utilizados en tests de fin de fabricación y en campo. Los resultados muestran que los métodos propuestos son un buen complemento del método tradicionalmente usado que consiste en observar el valor final del contenido de memoria. Además muestran que una adecuada combinación de estos métodos complementarios permite alcanzar casi los mismos valores de cobertura de fallas que se obtienen mediante la observación continua de todas las salidas del procesador, método comúnmente usado en tests de fin de fabricación, pero que usualmente no está disponible en campo. Un subproducto muy interesante de lo arriba expuesto es la descripción detallada del procedimiento para calcular la cobertura de fallas lograda mediante tests funcionales en campo por medio de un simulador de fallas convencional, una herramienta que usualmente se aplica en escenarios de test de fin de fabricación. Finalmente, otro resultado relevante en el área de test es un método para detectar fallas permanentes dentro de la lógica de coherencia de cache que está integrada en el controlador de cache de cada procesador en un sistema multi procesador. El método está basado en la ejecución de un programa de test en forma coordinada por parte de los diferentes procesadores. Por construcción, el método cubre completamente las fallas de la lógica mencionada
2019 | |
SISTEMAS ELECTRONICOS SOFTWARE-BASED SELF-TEST (SBST) |
|
Inglés | |
Universidad de la República | |
COLIBRI | |
https://hdl.handle.net/20.500.12008/22511 | |
Acceso abierto | |
Licencia Creative Commons Atribución - No Comercial - Sin Derivadas (CC - By-NC-ND 4.0) |
Sumario: | When electronic systems are used in safety critical applications, as in the space, avionic, automotive or biomedical areas, it is required to maintain a very low probability of failures due to faults of any kind. Standards and regulations play a significant role, forcing companies to devise and adopt solutions able to achieve predefined targets in terms of dependability. Different techniques can be used to reduce fault occurrence or to minimize the probability that those faults produce critical failures (e.g., by introducing redundancy). Unfortunately, most of these techniques have a severe impact on the cost of the resulting product and, in some cases, the probability of failures is too large anyway. Hence, a solution commonly used in several scenarios lies on periodically performing a test able to detect the occurrence of any fault before it produces a failure (in-field test). This solution is normally based on forcing the processor inside the Device Under Test to execute a properly written test program, which is able to activate possible faults and to make their effects visible in some observable locations. This approach is also called Software-Based Self-Test, or SBST. If compared with testing in an end of manufacturing scenario, in-field testing has strong limitations in terms of access to the system inputs and outputs because Design for Testability structures and testing equipment are usually not available. As a consequence there are reduced possibilities to activate the faults and to observe their effects. This reduced observability particularly affects the ability to detect performance faults, i.e. faults that modify the timing but not the final value of computations. This kind of faults are hard to detect by only observing the final content of predefined memory locations, that is the usual test result observation method used in-field. Initially, the present work was focused on fault tolerance techniques against transient faults induced by ionizing radiation, the so called Single Event Upsets (SEUs). The main contribution of this early stage of the thesis lies in the experimental validation of the feasibility of achieving a safe system by using an architecture that combines task-level redundancy with already available IP cores, thus minimizing the development time. Task execution is replicated and Memory Protection is used to guarantee that any SEU may affect one and only one of the replicas. A proof of concept implementation was developed and validated using fault injection. Results outline the effectiveness of the architecture, and the overhead analysis shows that the proposed architecture is effective in reducing the resource occupation with respect to N-modular redundancy, at an affordable cost in terms of application execution time. The main part of the thesis is focused on in-field software-based self-test of permanent faults. A set of observation methods exploiting existing or ad-hoc hardware is proposed, aimed at obtaining a better coverage, in particular of performance faults. An extensive quantitative evaluation of the proposed methods is presented, including a comparison with the observation methods traditionally used in end of manufacturing and in-field testing. Results show that the proposed methods are a good complement to the traditionally used final memory content observation. Moreover, they show that an adequate combination of these complementary methods allows for achieving nearly the same fault coverage achieved when continuously observing all the processor outputs, which is an observation method commonly used for production test but usually not available in-field. A very interesting by-product of what is described above is a detailed description of how to compute the fault coverage achieved by functional in-field tests using a conventional fault simulator, a tool that is usually applied in an end of manufacturing testing scenario. Finally, another relevant result in the testing area is a method to detect permanent faults inside the cache coherence logic integrated in each cache controller of a multi-core system, based on the concurrent execution of a test program by the different cores in a coordinated manner. By construction, the method achieves full fault coverage of the static faults in the addressed logic. |
---|