CSR: Small: Architectural Support for Programmable Memory Metadata

所有计算机都严重依赖巨大且不断增长的内存容量来处理日益多样化、具有挑战性和数据密集型的计算。为了有效和高效地利用存储器、提供高性能并保证安全和隐私，计算机系统将存储器元数据与存储在存储器中的所有信息相关联。此元数据存储有关数据的系统管理信息，而不是直接存储应用程序数据，这对正确性和性能至关重要。存储、访问和更新元数据在元数据成本和性能开销以及元数据可以提供的好处之间提供了独特的权衡机会。传统的方法是将元数据管理与内存本身的管理(以虚拟内存的形式)结合在一起，并且随着时间的推移发展得非常缓慢。然而，随着计算机存储系统向更大、更异质以及以更深层次组织的方向发展，这种方法正面临越来越多的困难。这种革命性的存储系统需要一个同样革命性的体系结构来管理和访问元数据。本项目正在研究和探索这样的元数据体系结构。这项研究非常必要，因为学术界和工业界都在不断寻找元数据的新用途，但往往留下了糟糕的元数据机制，这些机制的粒度错误，对性能的开销太高，而且太复杂，无法实现。因此，经常在性能、安全性、安全性和可靠性之间进行错误的权衡。这项研究的结果将直接导致安全、可靠、性能和高效的计算机系统，这些系统提供强大的隐私保障，这对科学家和整个社会所依赖的计算系统都是至关重要的。其方法是将元数据管理与数据内存管理(虚拟内存)解耦，同时在体系结构上公开元数据管理机制，以缓解上述许多限制。这种元数据体系结构将可从微体系结构内部编程，也可由系统软件编程，并提供更高水平的灵活性和能力。具体地说，研究回答了以下问题：(1)将通用内存元数据与系统关键转换和保护分离；(2)跨多个存储/访问机制协调元数据，在开销和粒度上进行不同的权衡；(3)为内存元数据构建可表达但高效的接口；以及(4)提供研究人员和行业分析内存管理想法所需的工具。

All computers heavily rely on large and growing memory capacity to handle increasingly diverse, challenging, and data-intensive computation. In order to utilize memory effectively and efficiently, provide high performance, and guarantee security and privacy, computer systems associate memory metadata with all information stored in memory. This metadata stores system-managed information about the data rather than application data directly and is vital for correctness and performance. Storing, accessing, and updating metadata presents unique tradeoff opportunities between metadata cost and performance overheads and the benefits metadata can provide. Conventional approaches have been to couple metadata management with the management of memory itself (in the form of virtual memory) and have evolved very slowly over time. However, this approach is facing increasing difficulties as computer memory systems evolve to be larger, more heterogeneous, and organized in deeper hierarchies. Such revolutionary memory systems require an equally revolutionary architecture for managing and accessing metadata. This project is researching and exploring such a metadata architecture. This research is sorely needed as both academia and industry are continually finding new uses for metadata but are often left with poor metadata mechanisms that are at the wrong granularity, have too high an overhead on performance, and are too complex to implement. Therefore, the wrong tradeoffs are often made between performance, security, safety, and reliability. The results of this research will directly lead to secure, reliable, performant, and efficient computer systems that provide strong privacy guarantees, which are all crucial to the computing systems both scientists and society in general have come to rely on. The approach is to decouple metadata management from data memory management (virtual memory) while exposing metadata management mechanism architecturally to alleviate many of the above limitations. This metadata architecture will be programmable from within the microarchiture and also by system software and provide new levels of flexibility and capabilities. Specifically, the research answers questions on how to: (1) decouple general memory metadata from system-critical translation and protection; (2) coordinate metadata across multiple storage/access mechanisms with different tradeoffs in overhead and granularity; (3) construct expressive yet efficient interfaces for memory metadata; and (4) provide the tools needed by researchers and industry to analyze memory management ideas.