Fault injection acceleration by architectural importance sampling

Abstract

Radiation-induced soft errors are major reliability concerns in advanced technology nodes. The de facto approach for evaluation of the soft error vulnerability is to perform a costly fault injection campaign. Due to the long residency of some errors in system states, the error has to be traced for even millions of cycles. However, only a very small portion of injected errors leads to the failure. This means that many simulation cycles are wasted as they contribute to no failure due to various masking effects. In this paper, we present an importance sampling technique based on Architecturally Correct Execution (ACE) analysis to identify the non-vulnerable time intervals in memory arrays and avoid unnecessary fault injections to speedup the soft error vulnerability evaluation process without sacrificing the accuracy. Our analysis reveals that this approach significantly expedites our architecture-level fault injection technique (on average by 13X).