Bit-Flip Aware Data Structures for Phase Change Memory

Big, non-volatile, byte-addressable, low-cost, and fast non-volatile memories like Phase Change Memory are appearing in the marketplace. They have the capability to unify both memory and storage and allow us to rethink the present memory hierarchy. An important draw-back to Phase Change Memory is limited write-endurance. In addition, Phase Change Memory shares with other Non-Volatile Random Access Memories an asym- metry in the energy costs of writes and reads. Best use of Non-Volatile Random Access Memories limits the number of times a Non-Volatile Random Access Memory cell changes contents, called a bit-flip. While the future of main memory is still unknown, we should already start to create data structures for them in order to shape the future era. This thesis investigates the creation of bit-flip aware data structures.The thesis first considers general ways in which a data structure can save bit- flips by smart overwrites and by using the exclusive-or of pointers. It then shows how a simple content dependent encoding can reduce bit-flips for web corpora. It then shows how to build hash based dictionary structures for Linear Hashing and Spiral Storage. Finally, the thesis presents Gray counters, close to bit-flip optimal counters that even enable age- based wear leveling with counters managed by the Non-Volatile Random Access Memories themselves instead of by the Operating Systems