A DATA AWARE APPROACH TO SALVAGE THE ENDURANCE OF PHASE-CHANGE MEMORY

摘要

Phase Change Memory (PCM) is an emerging non-volatile memory technology that could either replace or augment DRAM and NAND flash that are hindered by scalability challenges. PCM suffers from a limited endurance problem that needs to be alleviated before it can be endorsedudinto the memory stack. This thesis is based on the observation that the endurance problem and its ramification depend on the write data. Accordingly, a data-aware approach is applied to salvage the endurance of PCM at three different stages: pre-write fault avoidance, post-write fault tolerance and post-failure recovery.ududThe pre-write fault avoidance stage aims at reducing the endurance cost of servicing write requests. To this end, Cost Aware Flip Optimization (CAFO) is presented as an efficient technique to lessen the endurance degradation. Essentially, CAFO relies on a cost model that captures the endurance cost of programming memory cells based on their already stored values. Subsequently,the write data is encoded into a form that incurs a lower endurance cost than the original write data. Overall, CAFO is capable of reducing the endurance cost by up to 65% more than the existing schemes.ududWorn out PCM cells exhibit a stuck-at fault model which makes the manifestation of errors dependent on the values that cells are stuck at. When a write fails, the data is rewritten inverted. This dissertation shows that applying data inversion at the post-write fault tolerance stage exploits the data dependent nature of errors which enables ECCs to tolerate faults up to double their nominal capability. Furthermore, extensions to RDIS which is an ECC designed specifically for the stuck-at fault model are presented.ududAt the post-failure recovery stage, Data Dependent Sparing is presented to manage bad blocks in PCM. Departing from the observation that defective blocks in the context of the stuck-at fault model still exhibit a low write failure probability due to the data dependent nature of errors, this thesis takes the approach of reusing blocks that are defective yet better-than-bad through a dynamic management of the reserve spare space. Data Dependent Sparing is capable of increasing theudlifetime of PCM by up to 18%.ud