Neutron sensitivity and software hardening strategies for matrix multiplication and FFT on graphics processing units

Abstract

In this paper, we compare the radiation response of GPUs executing matrix multiplication and FFT algorithms. The provided experimental results demonstrate that for both algorithms, in the majority of cases, the output is affected by multiple errors. The architectural and code analysis highlight that multiple errors are caused by shared resources corruption or thread dependencies. The experimental data and analytical studies can be fruitfully employed to evaluate the expected error rate of GPUs in realistic applications and to design specific and optimized software-based hardening procedures

Exploiting non-constant safe memory in resilient algorithms and data structures

“…In this model, an adaptive adversary can corrupt up to δ memory cells of a large unreliable memory at any time (even simultaneously) during the execution of an algorithm. Resilient algorithmic techniques have been designed for many problems, including sorting [6], selection [7], dynamic programming [8], dictionaries [9], priority queues [10], matrix multiplication and FFT [11], K-d and suffix trees [12,13]. Resilient algorithms have also been experimentally evaluated [14,11,15,16].…”

“…Resilient algorithmic techniques have been designed for many problems, including sorting [6], selection [7], dynamic programming [8], dictionaries [9], priority queues [10], matrix multiplication and FFT [11], K-d and suffix trees [12,13]. Resilient algorithms have also been experimentally evaluated [14,11,15,16].…”