Input Size Effects on the Radiation-Sensitivity of Modern Parallel Processors

In this paper, we inspect the impact of modifying benchmarks' input sizes on parallel processors reliability. A larger input size imposes a higher scheduler strain, potentially increasing the parallel processor's radiation sensitivity. Additionally, input size affects the parallel codes throughput, the number of resources used for computation, and their criticality. The impact of input size is experimentally studied by comparing the radiation sensitivity of three modern parallel processors: Intel Xeon Phis, NVIDIA K20, and K40. Our test procedure has shown that parallel threads management significantly affects the device Silent Data Corruption sensitivity. Traditional reliability evaluation methodologies may result is a significant error (up to 200%) in the estimated sensitivity of parallel processors to radiation.