Ana Khorguani

Non-volatile memory (NVMM) technologies are a great opportunity to build fast fault-tolerant programs, as they provide persistent storage in main memory. However, since the processor caches remain volatile, solutions are needed to recover a consistent state from NVMM after a crash. In this thesis, we propose fast checkpointing approaches in NVMM to make multi-threaded programs fault tolerant. We focus on achieving high failure-free performance by flushing persistent data structures to NVMM periodically.

Our first work, ResPCT, considers a memory architecture where a single copy of the data is saved directly in NVMM. ResPCT uses In-Cache-Line logging to efficiently track modifications during failure-free execution, and to restore a consistent state after a crash. The ResPCT API enables programmers to position restart points in their program, which simplifies the identification of the persistent program state and can also help improving performance. Experiments with representative benchmarks and applications, show that ResPCT outperforms state-of-the-art solutions.

Our evaluation of ResPCT shows that relying on a single copy of the data in NVMM can limit performance. Therefore, in the second part of our study, we consider an alternative approach that integrates DRAM into the memory architecture. The proposed system involves maintaining a working copy of the data in DRAM, enabling the program to operate on this copy rather than on the slower version stored in NVMM. In this work, we compare several techniques to write data from DRAM to NVMM during checkpoints. Our results show that even though some techniques have advantages over others, choosing the most suitable method for transferring modifications from DRAM to NVMM depends on the specific characteristics of the applications.