Comparison of Fault Injection and Beam Experiments on GPUs

The goal of this proposal is to continue our evaluation of the resilience of modern parallel devices by comparing the experimentally measured FIT rate with the FIT rate estimated through fault injection. Our idea is to characterize the sensitivity of basic resources of GPUs (memories, functional units, conditional statements, mixed-precision hardware) through dedicated synthetic microbenchmarks. Then, we will measure the probability for faults in these basic resources through architectural level fault injectors (SASSIFI and NVBitFI). By multiplying the probability for a fault to happen in a basic resource with the amount of resources used for computation and the probability for a fault in each basic resource to affect computation we can have an estimation of the FIT rate of a code.