SHF:Small:Data Structures and Transactions for Emerging Nonvolatile Memory

无论是为了长期访问还是为了从系统崩溃中恢复，需要持久化的数据传统上都保存在磁盘和闪存驱动器上，这些驱动器非常慢。 在未来几年内，速度更快的持久内存预计将以合理的成本广泛使用，这增加了文件系统可能被简单的“始终可用”内存取代的可能性。该项目的智力价值在于解决这种“始终可用内存”的两个关键挑战：首先，确保系统崩溃后内存中的值始终相互一致，尽管传统缓存可能经常将数据无序地传递到内存;第二，尽管有缺陷的程序可能错误地修改任意存储器位置，但保护持久数据的结构完整性。 该项目更广泛的意义和重要性在于可编程性，可靠性和系统性能的显着改进，以及以比以前更低的成本承受停电和硬件故障的能力。 后一个好处可能是物联网中无处不在的传感器特别重要。该项目建立在首席研究员和同事之前的工作基础上，他们已经为持久性数据开发了正式的正确性标准，以及保证这种正确性的自动方法。 目前的项目正在进行三项主要的研究。 首先，它正在开发一个可重用的高性能持久数据结构库，特别强调利用高级语义来最大限度地减少插装开销，维护足够的信息来完成或撤销程序或系统崩溃后的部分操作，以及形式化和证明正确性。 其次，该项目正在开发将持久化操作组合成更大的原子事务的技术。 这项工作建立在过去的硬件和软件事务内存的经验，并包括非阻塞和基于锁的方法。 特别强调的是“提升”持久数据结构的操作，以便它们可以作为事务系统的可逆高级操作。 第三，该项目正在开发机制（包括用户级守护进程、基于编译器的沙箱、细粒度内存保护和虚拟化硬件的使用），以确保持久数据仅由受信任的库代码修改，从而在存在错误的应用程序时保护其结构完整性。

Data that need to be persistent, either for the sake of long-term access or for recovery from system crashes, have traditionally been kept on disk and flash drives, which are very slow. Within the next few years, much faster persistent memory is expected to be widely available at reasonable cost, raising the possibility that file systems might be replaced, at least in part, by memory that is simply "always available." The intellectual merits of the project lie in addressing two key challenges to such "always available memory": first, ensuring that the values in memory in the wake of a system crash are always mutually consistent, despite the possibility that traditional caches often pass data to the memory out of order; second, safeguarding the structural integrity of persistent data despite the possibility that buggy programs may erroneously modify arbitrary memory locations. The project's broader significance and importance lie in the promise of significant improvements in programmability, reliability, and system performance, and in the ability to survive power outages and hardware failures at much lower cost than has previously been possible. This latter benefit may be of particular importance for ubiquitous sensors in the Internet of Things.The project builds on prior work by the principal investigator and colleagues, which has developed formal correctness criteria for persistent data, together with automatic methods to guarantee this correctness. The current project is pursuing three major research thrusts. First, it is developing a library of reusable, high-performance persistent data structures, with particular emphasis on exploiting high-level semantics to minimize instrumentation overhead, maintaining sufficient information to complete or undo partial operations in the wake of a program or system crash, and formalizing and proving correctness. Second, the project is developing techniques to compose persistent operations into larger atomic transactions. This work builds on past experience with hardware and software transactional memory, and encompasses both nonblocking and lock-based approaches. Particular emphasis is being placed on "boosting" the operations of persistent data structures so that they can serve as reversible high-level operations of a transactional system. Third, the project is developing mechanisms (including user-level daemons, compiler-based sandboxing, fine-grain memory protection, and the use of virtualization hardware) to ensure that persistent data is modified only by trusted library code, thereby safeguarding its structural integrity in the presence of buggy applications.