CAREER: End-to-End Network Design for Unified Memory Disaggregation

现代云数据中心中的应用程序部署在资源容器中，以便彼此隔离。在这种容器化的数据中心中，内存滞留是一个普遍存在的问题，即使在其他机器中存在空闲内存，许多内存密集型应用程序也会陷入停顿。这将导致低利用率、内存碎片和总体成本的增加。在理论上，超高速网络上的内存分解可以将这些搁浅的内存聚集在一起，但使其实际应用面临着新的系统设计、算法和集成挑战。它们包括弥合本地内存访问与远程直接内存访问（RDMA）之间仍然相当大的延迟差距，透明地解决网络范围的容错、负载不平衡和性能隔离问题、可伸缩性，以及支持异构软件和硬件技术。本提案的总体研究目标是在超高速网络上实现统一分解内存（UDM）抽象，以一种快速、有弹性和可扩展的方式，将数据中心的存量内存作为可用内存池暴露给内存外容器，而无需对应用程序进行任何更改。通过设计一个全面的解决方案来解决上述挑战的主机级，网络级和端到端方面，本研究旨在使内存分解实用。具体来说，通过利用现代数据中心中内存密集型工作负载、超低延迟网络和多租户的独特特征，该提案将(i)设计一个低延迟主机网络堆栈；（ii）在整个网络中实现性能隔离；（iii）对网络范围内的不确定性（如故障和负载不平衡）提供弹性；（iv）整合对异构内存（例如，持久内存）、网络技术和资源管理软件的支持。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Applications in modern cloud datacenters are deployed in resource containers to isolate them from each other. Memory stranding is a pervasive problem in such containerized datacenters, where many memory-intensive applications grind to a halt even when free memory exists in other machines. This leads to low utilization, memory fragmentation, and overall increased cost. Memory disaggregation over ultra-fast networks can pool together such stranded memory in theory, but making it practical faces novel systems design, algorithmic, and integration challenges. They include bridging the still-sizable latency gap between local memory access vs. Remote Direct Memory Access (RDMA), transparently addressing network-wide fault-tolerance, load imbalance, and performance isolation issues, scalability, and enabling support for heterogeneous software and hardware technologies.The overarching research objective of this proposal is to realize a Unified Disaggregated Memory (UDM) abstraction over ultra-fast networks to expose stranded memory across the datacenter as a pool of available memory to out-of-memory containers in a fast, resilient, and scalable manner without any changes to the applications. By designing a comprehensive solution to address host-level, network-level, and end-to-end aspects of the aforementioned challenges, this research aims to make memory disaggregation practical. Specifically, by leveraging the unique characteristics of memory-intensive workloads, ultra-low-latency networks, and multi-tenancy in modern datacenters, this proposal will (i) design a low-latency host networking stack; (ii) enable performance isolation throughout the network; (iii) provide resilience to network-wide uncertainties such as failures and load imbalance; and (iv) incorporate support for heterogeneous memory (e.g., persistent memory), networking technologies, and resource management software.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.