Data Dependent Sparing to Manage Better-Than-Bad Blocks

We forecast that proper handling of unreliable storage blocks (e.g., "bad block management" in solid-state drives) will remain critical for future systems built with advanced and emerging memory technologies. This paper argues that the conventional block retirement and sparing approach--a block is retired as soon as it shows faulty behavior--is overly conservative and inefficient. We observe that it is highly unlikely that all faulty bits in a storage block manifest errors. Consequently, we propose data dependent sparing, a relaxed block retirement and sparing approach that recycles faulty storage blocks. At small management cost and with less than 1% sparing, data dependent sparing achieves the same lifetime as the conventional approach with 20% sparing.