NetPM: Co-designing Data Management and Networking Principles for Persistent Memory

在数据中心和面向互联网的系统中，以低成本提供高质量、可预测的服务的系统软件开销正成为问题，因为网络变得越来越快，存储设备也是如此。这些开销源于在不可靠的哑网络或存储硬件上实现各种网络或存储属性，例如数据组织的完整性、持久性、一致性和灵活性。实现这些属性会占用CPU，并导致高速网络和存储设备（包括新兴的持久性内存（PM））的利用率不足。这一观察结果意味着运营商需要投入更多的计算硬件资源，以确保实现所需的服务水平目标。更糟糕的是，尽管实现存储属性的成本很高，但PM设备中可用的硬件卸载机会比块设备少，块设备充当I/O外围设备。该项目旨在通过重新利用网络堆栈或协议中可用的功能来解决这个问题，这些功能通常可以卸载到网络硬件，以实现存储属性，从而使存储堆栈可以最大限度地减少其操作。这里的关键见解是，根据定义，存储堆栈所需的许多属性也可用于网络堆栈或端到端协议（如TCP）中，并且这些属性可以被重新利用，特别是当存储数据不是驻留在块设备中，而是驻留在新兴的字节可寻址PM设备中时，其数据结构类似于内存中的数据结构。这种体系结构很难实现。由于网络和存储堆栈是隔离设计的，现有的数据结构在这些堆栈之间是不可重用的，这意味着我必须设计满足网络和存储需求的新数据结构。例如，由于当前的网络数据结构是短暂的，它们既不是命名的，也不是崩溃一致的。此外，由于当前的标准API和编程模型是为网络和存储访问的单独操作而设计的，因此我必须设计新的、有用的编程抽象，以有效地利用网络辅助存储堆栈，然后在这些抽象之上构建具体的网络存储应用程序。因此，为了实现和验证我的方法，我必须设计整个软件栈，其中包括低级别的数据包I/O框架，持久网络数据结构，高级API和应用程序。该项目的计划开始通过启用技术来重新利用网络功能来实现存储属性，包括设计数据结构，算法和API。然后，它将为网络存储系统设计和实现具体的软件栈，包括文件系统、键值存储和事务数据库，其中每一个都有不同的要求和目标工作负载。除了基于实际工作负载对这些新系统进行实验评估外，该项目还将通过构建更复杂的应用程序来验证结果，例如针对图结构数据的分布式机器学习。我在最相关的领域领导研究工作的记录已经得到证明，这是网络和操作系统之间的交叉点，我相信我有一个雄心勃勃的目标，有一个合理的工作和资源计划和合作伙伴的支持。

System software overheads are becoming problematic in both data centre and Internet-facing systems to offer high quality, predictable services at low costs, because the networks are becoming faster and faster, and so are storage devices. These overheads stem from achieving various networking or storage properties, such as integrity, durability, consistency and flexibility of data organization, over unreliable, dumb networks or storage hardware. Achieving these properties overwhelms the CPUs and results in underutilization of high-speed networks and storage devices, including emerging persistent memory (PM). This observation implies that the operators need to invest more compute hardware resources to ensure the desired service-level objectives. Even worse, although achieving the storage properties is costly, fewer opportunities of hardware offloading are available in PM devices than in block devices, which act as I/O peripherals. This project aims to address this problem by repurposing features available in the network stack or protocols, which often can be offloaded to the networking hardware, for achieving storage properties, so that the storage stacks can minimise their operations. The key insight here is that many properties required by the storage stack are, by definition, also available in the network stack or end-to-end protocols like TCP, and these properties could be repurposed particularly when storage data is resident in not the block devices, but the emerging byte-addressable PM devices whose data structures resemble in-memory data structures.Despite the simple high-level concept, this architecture is difficult to achieve. Since network and storage stacks have been designed in isolation, existing data structures are not reusable between these stacks, meaning that I must design new ones that meet both networking and storage requirements. For example, since the current networking data structures are ephemeral, they are neither named nor crash consistent. Further, since the current standard APIs and programming models are designed for separate operations for networking and storage access, I must design new, useful programming abstractions that effectively utilize the network-aided storage stack and then build concrete networked storage applications on top of these abstractions. Therefore, to achieve and verify my approach, I must design the whole software stack across a variety of components that include low-level packet I/O framework, persistent networking data structures, high-level APIs and the applications.The programme of the project sets off by enabling techniques to repurpose the networking features for implementing the storage properties, including designing data structures, algorithms and APIs. It will then design and implement concrete software stacks for networked storage systems, including file system, key-value store and transactional database, each of which has different requirements and target workloads. In addition to experimental evaluation of those new systems based on realistic workloads, the project will finally validate the results by building more complex applications, such as distributed machine learning for graph structured data.I have a proven track record of leading the research work in the most relevant area, which is an intersection between networking and OS, and believe I have an ambitious goal backed by a reasonable work and resource plan and partner.