Pasar al contenido principal

Resultados relacionados

Resultados totales (Incluyendo duplicados): 4
Encontrada(s) 1 página(s)

RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Tesis de maestría (MasterThesis)

Toolchain automatizado para la integración de Kernels CUTLASS en motores TensorRT, Automated Toolchain for the Integration of CUTLASS Kernels into TensorRT Engines, Toolchain automatitzat per a la integració de Kernels CUTLASS en motors TensorRT

  • Ripoll-Acosta, Alejandro
[ES] El despliegue de Redes Neuronales Convolucionales (CNN) en sistemas críticos exige tanto un alto rendimiento como estrictas garantías de seguridad funcional frente a fallos físicos de hardware. Sin embargo, motores de inferencia de referencia como NVIDIA TensorRT operan bajo un paradigma de «caja negra» que impide la inyección de mecanismos de tolerancia a fallos a nivel de kernel. Este Trabajo de Fin de Máster presenta el diseño, implementación y evaluación de un toolchain automatizado que sustituye las convolu- ciones nativas de TensorRT por kernels altamente optimizados generados con CUTLASS. Esta transición hacia un paradigma de «caja blanca» otorga un control granular sobre la ejecución, permitiendo la integración de un mecanismo de saneamiento in-situ. Dicho mecanismo intercepta activaciones anómalas mediante la inyección de una operación de recorte asimétrico directamente en el epílogo del kernel, acotando los valores corruptos antes de que colapsen el grafo tensorial. La evaluación experimental, realizada sobre un modelo YOLOv8n en un dispositivo NVIDIA Jetson AGX Orin, demuestra que el toolchain propuesto ofrece un rendimiento competitivo frente al motor propietario, siendo capaz de generar un engine equivalente compatible con el motor de inferencia de NVIDIA. Sometido a una severa campaña de inyección de fallos (bit-flips), el mecanismo de protección logra erradicar por completo los colapsos del sistema (Tasa de Fallo Fatal del 0,00 %) garantizando la continuidad de la inferencia con una sobrecarga computacional inferior al 1 %.
RUA. Repositorio Institucional de la Universidad de Alicante
Publicación Otros (Other)

Exploring the limits of software-only diverse redundancy on an Intel GPU

  • Andriotis, Nikolaos
  • Serrano-Cases, Alejandro
  • Alcaide, Sergi
  • Cazorla, Francisco J.
  • Abella, Jaume
  • Serrano-Cases, Alejandro|||/items/09bc6bdc-e4fd-4043-835a-48fee62ed569
The advent of Autonomous Driving (AD) in the automotive domain, and AD’s high dependence on compute-intensive Artificial Intelligence (AI) based tasks makes the use of graphics processing units (GPUs) mandatory in those systems. However, AD controls systems with the most stringent safety integrity levels like steering and braking that impose the use of diverse redundancy in accordance with ISO 26262 functional safety standard for road vehicles. Since commercial GPUs lack explicit support for diverse redundancy, software-only solutions making redundant tasks use disjoint sets of computing resources should be explored. So far, some solutions have been proposed to achieve high (yet not full) degrees of diversity on NVIDIA GPUs, as well as full diversity (for the computing resources) on Intel GPUs subject to meeting two constraints: (1) being able to split the work into a fixed number of software threads matching the number of hardware threads, and (2) having a hardware scheduler allocating all hardware threads in a perfect round-robin fashion. The first constraint cannot be easily met and may harm performance due to work imbalance, whereas the second is fully dependent on the GPU implementation, and can only be assessed empirically to some extent. This paper provides a new solution to achieve diverse redundancy for the computing resources in COTS GPUs, proven on an Intel X GPU, overcoming the limitations of existing solutions. In particular, our approach avoids constraints on the number of software threads used and the hardware scheduler allocation policies. Our approach builds on a first phase making each software thread select the piece of work to carry out based on the hardware thread where it has been scheduled with no need for synchronization, and a second phase in which the limitations of the hardware scheduling policy are managed securing workload execution while sacrificing very little diversity. We realize our approach in a last generation Intel GPU for which we show low performance overheads, but the very same strategy can be ported to other GPU families (e.g., NVIDIA, and Arm).
UPCommons. Portal del coneixement obert de la UPC
Publicación Artículo científico (JournalArticle)

Supporting timing-related metrics for autonomous driving frameworks in CyberRT

  • Alcón Doganoc, Miguel
  • Mezzetti, Enrico|||0000-0002-1886-2931
  • Abella Ferrer, Jaume|||0000-0001-7951-4028
  • Cazorla Almeida, Francisco Javier
The provision of increasingly advanced autonomous software functionalities builds on cutting-edge autonomous driving frameworks to enable modular interactions among multiple software components. This approach helps to support functional cause-effect chains from multiple sensors to actuators. The complexity of the (software) component interactions makes it more difficult to ascertain the correctness of the timing behavior of the system. This is so because traditional timing-related metrics like worst-case execution and worst-case response time do not capture the inter-dependency in cause-effect chains between the input sampling time and the time at which computation based on those inputs is performed. Complementary timing-related metrics, such as maximum reaction time and maximum data age have been considered to capture timing requirements, typically with an end-to-end scope, in cause-effect chains. These metrics have been formalized and demonstrated in ROS2-based automotive and autonomous driving setups. However, the formalization of those metrics, which is necessary for deriving analytical lower and upper bounds and monitoring them at run-time, largely depends on the execution model and semantics offered by the run-time. Any concrete application of those metrics need to be tailored and adapted to the system at hand. Apollo auto is a popular, industrial-quality, open-source autonomous driving framework that is seeing increasing adoption both for industrial and academic projects. Apollo builds on CyberRT, an ad-hoc run-time that is similar in mechanism and intent to ROS2 but differentiates from it with respect to execution model and supported semantics. In contrast to ROS2, CyberRT is highly specialized to support the Apollo AD framework, is neither extensively documented or thoroughly analysed in the literature, especially in relation to execution model and instantiation of timing-related metrics. In this work, for the first time, we provide an insightful analysis and discussion on CyberRT execution model and semantics, starting from its raw and non-extensively documented codebase. Based on the identified semantics, we elaborate a formalization of timing-related metrics on CyberRT, across different granularity scopes, namely end-to-end and node levels. In particular, we develop on the importance of node-level timing properties to intercept any latent timing misbehavior before it is too late, and it severely impacts end-to-end execution. We provide a concrete mapping of a comprehensive set of timing-related metrics to the CyberRT execution model, both at end-to-end and node level, and develop a monitoring library that allows to intercept them on the specific software stack. We exploit the proposed library on a set of Apollo autonomous driving scenarios to demonstrate its effectiveness in monitoring the considered timing metrics and to promptly intercept a subtle timing misbehavior beyond end-to-end execution scope in a representative autonomous driving stack., T his work has been supported by the European Union’s Horizon Europe research and innovation programme under grant agreement No 101139892 (EdgeAI-Trust); by the project PLEC2023-010240, funded by MICIU/AEI /10.13039/501100011033 (CAPSUL-IA);andbytheMiguelAlcónfellowshipwithinthe”GeneraciónD”initiative,Red.es,Ministerio parala Transformación Digital ydelaFunciónPública,fortalentattraction(C005/24-EDCV1),fundedbytheEuropeanUnionNextGenerationEU funds through PRTR., Peer Reviewed
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

Supporting timing-related metrics for autonomous driving frameworks in CyberRT

  • Alcón Doganoc, Miguel
  • Mezzetti, Enrico
  • Abella Ferrer, Jaume
  • Cazorla Almeida, Francisco Javier
The provision of increasingly advanced autonomous software functionalities builds on cutting-edge autonomous driving frameworks to enable modular interactions among multiple software components. This approach helps to support functional cause-effect chains from multiple sensors to actuators. The complexity of the (software) component interactions makes it more difficult to ascertain the correctness of the timing behavior of the system. This is so because traditional timing-related metrics like worst-case execution and worst-case response time do not capture the inter-dependency in cause-effect chains between the input sampling time and the time at which computation based on those inputs is performed. Complementary timing-related metrics, such as maximum reaction time and maximum data age have been considered to capture timing requirements, typically with an end-to-end scope, in cause-effect chains. These metrics have been formalized and demonstrated in ROS2-based automotive and autonomous driving setups. However, the formalization of those metrics, which is necessary for deriving analytical lower and upper bounds and monitoring them at run-time, largely depends on the execution model and semantics offered by the run-time. Any concrete application of those metrics need to be tailored and adapted to the system at hand. Apollo auto is a popular, industrial-quality, open-source autonomous driving framework that is seeing increasing adoption both for industrial and academic projects. Apollo builds on CyberRT, an ad-hoc run-time that is similar in mechanism and intent to ROS2 but differentiates from it with respect to execution model and supported semantics. In contrast to ROS2, CyberRT is highly specialized to support the Apollo AD framework, is neither extensively documented or thoroughly analysed in the literature, especially in relation to execution model and instantiation of timing-related metrics. In this work, for the first time, we provide an insightful analysis and discussion on CyberRT execution model and semantics, starting from its raw and non-extensively documented codebase. Based on the identified semantics, we elaborate a formalization of timing-related metrics on CyberRT, across different granularity scopes, namely end-to-end and node levels. In particular, we develop on the importance of node-level timing properties to intercept any latent timing misbehavior before it is too late, and it severely impacts end-to-end execution. We provide a concrete mapping of a comprehensive set of timing-related metrics to the CyberRT execution model, both at end-to-end and node level, and develop a monitoring library that allows to intercept them on the specific software stack. We exploit the proposed library on a set of Apollo autonomous driving scenarios to demonstrate its effectiveness in monitoring the considered timing metrics and to promptly intercept a subtle timing misbehavior beyond end-to-end execution scope in a representative autonomous driving stack., T his work has been supported by the European Union’s Horizon Europe research and innovation programme under grant agreement No 101139892 (EdgeAI-Trust); by the project PLEC2023-010240, funded by MICIU/AEI /10.13039/501100011033 (CAPSUL-IA);andbytheMiguelAlcónfellowshipwithinthe”GeneraciónD”initiative,Red.es,Ministerio parala Transformación Digital ydelaFunciónPública,fortalentattraction(C005/24-EDCV1),fundedbytheEuropeanUnionNextGenerationEU funds through PRTR., Peer Reviewed, Postprint (author's final draft)