Does XORing Pointers Save Bitflips for NVRAM?

Abstract

New classes of non-volatile random access memory technology have the potential to unify the memory and the storage layers of modern computers, both for end-users and for data centers. The emerging non-volatile memory technologies have (or are expected to have) a speed close to the speed of current DRAM and are as DRAM byte-addressable, but they have the capacity and low cost of storage technologies. Unlike DRAM, they are non-volatile and they draw energy only for active reading and writing. Some, such as Phase Change Memory, also have a limited resilience, as each overwrite of a cell content can damage the cell removing its ability to store a bit faithfully. In a scenario where a non-volatile RAM replaces the current memory, bit-flip pressure becomes an important criteria for operating system and data structure design. As observed by Bittman and colleagues, replacing traditional data structures with ones where pointers are replaced by the exclusive-or of pointers can save bit-flips. We investigate this observation further and show that this is not the case for very large data structures or for data structures modified over long period of times. We then draw conclusions about virtual memory management in future operating systems optimized for the use of non-volatile random access memories.