Hybrid Hardware/Software Detection of Multi-Bit Upsets in Memory

Bit flips in main memory can be caused by a multitude of environmental effects, such as heat or radiation, as well as by malicious actors exploiting Rowhammer-style hardware vulnerabilities. The industry-standard countermeasure is SEC-DED ECC memory, which can reliably correct single- and detect double-bit flips in a data word. However, larger multi-bit upsets (MBUs) regularly occur in real-world systems, and - as shown by an analysis in this paper - have a high probability of being miscorrected. Software-implemented hardware fault tolerance (SIHFT) mechanisms can flexibly handle MBUs, but incur significant runtime costs. In this paper, we propose to combine hardware ECC as a low-cost detector and SIHFT as a handler for miscorrected MBUs that recategorizes them as uncorrectable. A preliminary evaluation on the basis of differential checksums shows a 98.5% reduction in miscorrected silent data corruptions with a very moderate execution-time overhead.