Collaborative Research: PPoSS: Planning: Towards an Integrated, Full-stack System for Memory-centric Computing

随着当今系统处理的数据量持续增加，存储系统的传统组织结构正在转变，以适应这种快速增长。以数据为中心的应用，如不规则图挖掘算法、分布式机器学习和基因组测序，需要大量的数据来计算和存储，并生成大量的中间数据来在计算资源中移动。以内存为中心的计算是克服当前系统性能瓶颈的潜在解决方案。内存附近或内存中的计算可以通过减少内存和主机处理单元之间的数据移动来缓解带宽限制；具有由所有处理单元共享的快速互连的远程内存池可以克服当前的容量限制。这两种解决方案都有望打破记忆墙。然而，通过直接集成来释放这两种解决方案的能力是具有挑战性的。在这个项目中，研究人员提出了一个集成的全栈系统来实现以内存为中心的计算(SMC2)。该系统将结合新兴的近内存数据处理器(NDP)和可扩展的远程内存池，以最大限度地减少图挖掘应用程序中内存访问的性能影响。研究任务包括体系结构、软件/硬件接口、编程模型/编译器和性能模型/优化方面的优化。首先，重新访问该体系结构，以利用NDP硬件构建支持智能数据预取和投机性数据推送的主动存储系统。接下来，重新设计系统软件以支持NDP函数调用、数据推送操作和虚拟化。然后，通过新的系统抽象，提出了一种新的编程模型，允许程序员指定哪些任务可以在NDP资源上运行，并支持有效的NDP到NDP通信。最后，提出了一种新的系统性能模型和优化框架。通过将这四个部分结合在一起，建议的系统支持可以通过新的系统抽象和理论最大化以内存为中心的计算的性能。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

As the volume of data being processed by today’s systems continues to increase, the traditional organization of memory systems is shifting to accommodate that accelerating growth. Data-centric applications such as irregular graph-mining algorithms, distributed machine learning, and genome sequencing require a large amount of data to compute and store, and generate massive amounts of intermediate data to move around the compute resources. Memory-centric computing is a potential solution to overcome the performance bottleneck of current systems. Near or in-memory computing can mitigate the bandwidth limitations with fewer data movements between the memory and host processing units; a remote memory pool with a fast interconnect shared by all processing units can overcome the current capacity constraints. Both solutions are promising for breaking down the memory wall. However, it is challenging to release the power of both solutions with direct integration. In this project, the investigators propose an integrated, full-stack system to enable memory-centric computing (SMC2). The system will incorporate the emerging near-memory data processors (NDP) and an extensible remote memory pool to minimize the performance impact of memory accesses in graph-mining applications. The research tasks include optimizations in architecture, the software/hardware interface, programming models/compilers, and performance models/optimization. First, the architecture is revisited to utilize the NDP hardware to build an active memory system that supports intelligent data prefetch and speculative data push. Next, the system software is redesigned to support NDP function calls, data-push operations, and virtualization. Then, with new system abstractions, a new programming model is proposed to allow programmers to specify which tasks can run on the NDP resources, and to support efficient NDP-to-NDP communication. Lastly, a new system performance model and optimization framework are incorporated. By putting the four pieces together, the proposed system support can maximize the performance of memory-centric computing with new system abstractions and theories.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.