Saving Bit-flips through Smart Overwrites in NVRAM

Abstract

New generations of non-volatile random access memories will combine the best features of memory (access times, byte addressability) with the best features of storage (non-volatility, low costs per byte). Some, like PCM, have a limited endurance. All will only consume energy when accessed, but writes will use much more energy than reads. These characteristics put a cost on flipping bits in memory. Bit-flip aware data structures lower the number of bits flipped by not resetting fields to zero to indicate a deleted record but by using bit-maps. If given a choice of where to over-write data, they will select the location which results in a lower number of bit-flips. We calculate the expected bit-flip savings of this strategy and derive a rule to determine the number of the possible candidate locations.