Reducing the Code Degree Of Parallelism to Increase GPUs Reliability
UFRGS, Universidade Federal do Rio Grande do Sul, Porto Alegre, Brazil
The 10th IEEE Workshop on Silicon Errors in Logic – System Effects (SELSE-10), 2014
@article{rech2014reducing,
title={Reducing the Code Degree Of Parallelism to Increase GPUs Reliability},
author={Rech, P and Carro, L},
year={2014}
}
A higher Degree of Parallelism decreases the code execution time. However, to manage the increased number of parallel processes a higher scheduling strain is required and caches, registers, and other resources utilization will be affected. All these parallelism management variations may have the countermeasure of increasing the GPU neutron sensitivity. The results of an extensive neutron radiation test on the Degree of Parallelism effects on GPUs reliability confirm this hypothesis. Finally, the Mean Executions Between Failures metric is used to evaluate the number of executions computed correctly by the GPU on a practical application.
July 3, 2014 by hgpu
