Evaluating the soft error sensitivity of a GPU-based SoC for matrix multiplication

Abstract

System-on-Chip (SoC) devices can be composed of low-power multicore processors combined with a small graphics accelerator (or GPU) which offers a trade-off between computational capacity and low-power consumption. In this work we use the LLFI-GPU fault injection tool on one of these devices to compare the sensitivity to soft errors of two different CUDA versions of matrix multiplication benchmark. Specifically, we perform fault injection campaigns on a Jetson TK1 development kit, a board equipped with a SoC including an NVIDIA "Kepler" Graphics Processing Unit (GPU). We evaluate the effect of modifying the size of the problem and also the thread-block size on the behaviour of the algorithms. Our results show that the block version of the matrix multiplication benchmark that leverages the shared memory of the GPU is not only faster than the element-wise version, but it is also much more resilient to soft errors. We also use the cuda-gdb debugger to analyze the main causes of the crashes in the code due to soft errors. Our experiments show that most of the errors are due to accesses to invalid positions of the different memories of the GPU, which causes that the block version suffers a higher percentage of this kind of errors.

Artificial Neural Networks for Space and Safety-Critical Applications: Reliability Issues and Potential Solutions

“…Since the injection is purely software, the execution can be performed on the live hardware, significantly reducing the injection overhead (one fault injection takes as much time as a normal execution) and the fault injection design (no need to simulate and control the circuit). The ease of use and efficiency of software fault injection makes it popular among research groups evaluating the reliability of neural networks [62], [69], [70], [101], [161]- [163]. However, the very high level of abstraction at which software fault injection is performed imposes a need to carefully select the error model.…”

Comparison of Parallel Implementation Strategies in GPU-Accelerated System-on-Chip Under Proton Irradiation

“…An alternative method to analyze the vulnerability of different components of the GPU is to inject faults at the architecture-level or into compiler-level intermediate representations [38], [39]. In [40] we employed the LLFI injection tool to evaluate the soft error sensitivity of the algorithms Elem and Block in the same Jeston TK1 used in this paper. We injected single bit-flips in the results of the instructions of randomly chosen threads.…”

Experimental Findings on the Sources of Detected Unrecoverable Errors in GPUs

“…Previous studies have stated that parallel architectures, particularly GPUs, have a high fault rate because of the high amount of available resources [7], [12], [13]. Recent works have identified some peculiar reliability weaknesses of GPUs architecture, suspecting that the corruption of the GPU hardware scheduler or shared memories can severely impact the computation of several parallel threads [3], [7], [12], [14], [15]. As a result, multiple GPU output elements can potentially be corrupted, effectively undermining several applications' reliability, including CNNs [16], [17].…”

“…Few works have studied the DUEs on GPUs on beam experiments [3], [10], [18]- [20]. However, most of the mentioned works do not present a detailed analysis of the events that cause DUEs on GPUs, not allowing a deep investigation of the weakness of the GPUs.…”

“…Most recent studies target Silent Data Corruption (SDC) in their evaluation. SDCs, being undetectable, are, in fact, considered the main threat for modern computing device's reliability [2], [3]. Detected Unrecoverable Errors (DUEs), such as device hangs, application crashes, or functional interruptions, are considered less harmful as, being detectable by definition, they could be easily handled using solutions such as checkpoints and software/hardware watchdogs [4]- [6].…”