Exploiting Narrow Data-Width to Mask Soft Errors in Register Files

摘要

The dependability of computing, caused by soft errors, has become a growing design concern in the safety critical systems. Since Register Files (RFs) are very frequently accessed and errors occurred in them will propagate to other components quickly, RFs are among the major reasons for affecting systemic reliability. Current protecting techniques usually provoke significant power penalty and performance degradation. This paper proposes a lightweight software implemented method for mitigating soft errors in RFs. Based on the observation of many narrow data-width of registers' value, which indicates a large fraction of unused bits of register data, the masking operations are inserted to clear the possible errors in these bits for reducing the window of vulnerability for RFs. To improve the effectiveness, the effect of each masking range is calculated, and the covered masks analysis can remove the unnecessary masks without scarifying the errors coverage. Under the user-defined overhead constrain, the most cost-effective masking operations can be automatically selected. Experimental results from several benchmarks indicate that the reliability of programs have been averagely improved for 16.8% with only 3.3% performance overhead.