学术文献库

Non-Volatile Memory devices may soon be a part of main memory, and programming models that give programmers direct access to persistent memory through loads and stores are sought to maximize the performance benefits of these new devices. Direct access introduces new challenges. In this work, we identify an important aspect of programming for persistent memory: the persistent data retention model. A Persistent Data Retention Model describes what happens to persistent data when code that uses it is modified. We identify two models present in prior work but not described as such, the Reset and Manual Model, and we propose a new one called the Automatic Model. The Reset model discards all persistent data when a program changes leading to performance overheads and write amplification. In contrast, if data is to be retained, the Manual Model relies on the programmer to implement code that upgrades data from one version of the program to the next. This reduces overheads but places a larger burden on the programmer. We propose the Automatic Model to assist a programmer by automating some or all of the conversion. We describe one such automatic approach, Lazily Extendible Data Structures, that uses language extensions and compiler support to reduce the effort and complexity associated with updating persistent data. We evaluate our PDRMs in the context of the Persistent Memory Development Kit (PMDK) using kernels and the TPC-C application. Manual Model shows an overhead of 2.90% to 4.10% on average, and LEDS shows overhead of 0.45% to 10.27% on average, depending on the workload. LEDS reduces the number of writes by 26.36% compared to Manual Model. Furthermore, LEDS significantly reduces the programming complexity by relying on the compiler to migrate persistent data.