Using Memory-style Storage to Support Fault Tolerance in Data Centers

Next-generation nonvolatile memories combine byteaddressability and high performance of memory with nonvolatility of disk/flash. They promise emerging memory-style storage (MSS) systems that are directly attached to the memory bus, offering fast load/store access and data persistence in a single level of storage. MSS can be especially attractive in data centers, where fault tolerance support through storage systems is critical to performance and energy. Yet existing fault tolerance mechanisms, such as logging and checkpointing, are designed for slow block-level storage interfaces; their design choices are not wholly suitable for MSS. The goal of this work is to explore efficient fault tolerance techniques that exploit the fast memory interface and the nature of single-level storage. Our priliminary exploration shows that, by reducing data duplication and increasing application parallelism, such techniques can substantially improve system performance and energy consumption.