Transitioning scientific applications to using non-volatile memory for resilience

Scientific applications often run for long periods of time, and as a result, frequently save their internal states to storage media in cases of unexpected interruptions (e.g., hardware failures). Emerging non-volatile memory (NVRAM) can write up to 40× faster than traditional mechanical storage devices, providing an attractive medium for this purpose. This paper investigates the implications of transitioning a scientific application, Fluidanimate, to use NVRAM for fault tolerance. In particular, we evaluate the performance implications and ease-of-use of four fault-tolerance approaches: 1) logging through transactions, 2) multi-versioning through copy-on-write operations, and 3) checkpointing through IO operations (e.g., fwrite) on a direct access (DAX) filesystem and 4) checkpointing with a DRAM cache. Our study results in three key findings. First, additional changes to the application are required to take advantage of the increase in IO speed provided by NVRAM. Second, the performance scalability of the approaches lack when considering a single process. Third, NVRAM can increase reliability in a distributed computing environment by allowing individual nodes to error and automatically recover before the rest of the system notices.