XPS: EXPL: FP: Architecture and Software for Scalable Persistent Memory

新兴的存储类内存(SCM)技术将硬盘和固态硬盘(SSD)等存储设备的非易失性与以字节粒度而不是以块级别访问数据的能力结合在一起。应用程序开发人员可以专注于单个细粒度的存储抽象，而不是处理DRAM内存的字粒度访问和基于文件的磁盘存储的数据块接口。直接从SCM访问数据而不是软件仲裁的块访问具有显著的性能优势，特别是对于基于图形和关系处理的现代数据密集型应用程序。然而，在出现意外的机器故障和软件中止的情况下，使用直接SCM访问会给应用程序带来很大的负担，以实现事务性和健壮的执行。这项研究将确定用于开发健壮的基于SCM的软件的轻量级、可扩展的体系结构和软件技术。其目标是通过将程序员从管理事务原子性和持久性的复杂性中解放出来，使持久存储器的软件开发可与对传统DRAM主存储器的编程相媲美。将开发基于处理器和内存控制器体系结构增强的硬件解决方案，以及轻量级持久化库形式的软件方法。这些技术将被扩展以处理多个线程和核心，并向外扩展到多个CPU插槽和分布式集群体系结构。

Emerging Storage Class Memory (SCM) technologies combine the non-volatility of storage devices like hard disks and solid state drives (SSDs) with the ability to access data at byte granularity rather than at a block level. Application developers can focus on a single fine-grained storage abstraction, rather than deal with word-granularity access for DRAM memory and a block interface for file-based disk storage. Accessing data directly from SCM instead of software-arbitrated block access has significant performance advantages, especially for modern data intensive applications based on graph and relation processing. However the use of direct SCM access places a significant burden on the application to achieve transactional and robust execution in the presence of unexpected machines failures and software aborts. The research will identify lightweight, scalable architectural and software techniques for developing robust SCM-based software. The goal is to make the development of software for persistent memory comparable to programming conventional DRAM main-memory by relieving the programmer from the complexities of managing transactional atomicity and persistence. Both hardware solutions based on enhancements to the processor and memory-controller architecture, and software approaches in the form of a lightweight persistence library will be developed. The techniques will be scaled up to handle multiple threads and cores, and scale out to multiple CPU sockets and to distributed clustered architectures.