Year: 2021
Type: dissertation
Abstract: Currently, Graphics Processing Units (GPUs) are crucial devices able to boost the execution of complex algorithms in the scientific and artificial intelligence domains.Moreover, GPU-based platforms are also relevant components now included in several safety-critical applications (e.g., in the automotive and autonomous machines fields), where reliability and functional safety are essential requirements.The doctoral research activities were focused on identifying new online techniques for testing and mitigation of faults affecting GPUs.This work also describes a new microarchitectural GPU model (FlexGripPlus), a supporting tool for detailed reliability evaluation.FlexGripPlus can also be used to support the development of functional test approaches and mitigation solutions.This GPU model is compatible with the CUDA programming environment and is a corrected and extended version of a previous GPU model implementing the G80 architecture by NVIDIA.The GPU model's extensions include support for floating-point operations and for the execution of trigonometric and transcendental operations using Special Function Units.FlexGripPlus was used to develop, evaluate, and validate functional test techniques based on the software-based self-test (SBST) strategy.More in detail, two main strategies are proposed for GPUs: i) the multi-kernel test approach, and ii) the modular approach of testing.On the one hand, the multi-kernel approach exploits the GPU's main operative behavior as a special-purpose accelerator.This approach is based on the parallel execution of test programs intended to detect and propagate the fault effects on any GPU core's available outputs.For this purpose, a set of parallel test programs with different GPU configuration parameters allows the generation of test patterns after dividing a target module into fault groups.This division process allows the coverage of sensitive locations that implicitly remain constant by the effect of the configuration of a parallel program.Moreover, this technique also uses a threadbased method to propagate and identify faults into the GPU's available outputs, simplifying the evaluation of specific modules in the GPU architecture, such as the pipeline registers.On the other hand, the modular testing approach combines the microarchitectural details of a given module, its functional operation, its major constraints, and iii • To my Mom and Dad, who encouraged me with tremendous moral support and reminded me to follow my dreams.• To my sister Melissa and my cousin Sebastian, who supported me with good vibes and fun when were required.• I would like to express my respect and huge thanks to my mentor and advisor, Prof. Matteo Sonza Reorda, for the vast support and for showing me how simple ideas can turn into reality with discipline and hard work.• To my friends Aleksa D., Andrea F. and Pratyush K. for the academic and moral support in all steps of this journey.• To my friends "ESRs" from RESCUE-ETN project for the huge support; each meeting was a way to recharge myself with good energy to continue working with more determination.• To the RESCUE-ETN
Source: PORTO@Iris
Authors: Josie E. Rodriguez Condia, Josie Esteban
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Keywords: graphics processing units, online assessment, fault mitigation, software-based self-test, modular testing, fault propagation, pipeline registers, safety-critical systems, functional safety, reliability analysis, CUDA, floating-point operations, test pattern generation, fault detection, autonomous vehicles
Open Access status: green
OpenAlex ID: W3203861171
