Memory Management Issues in Large Scale Shared Memory MIMD Multiprocessors

在可以扩展到大量处理器的共享存储器设计中，存储器访问的非一致性几乎是存储器体系结构的一个不可避免的特征，从而产生了日益重要的NUMA(非均匀存储器访问时间)多处理器。管理代码和数据的放置和移动对NUMA体系结构中的性能至关重要。这被称为NUMA问题。并行体系结构还加剧了I/O瓶颈问题，这种体系结构拉大了处理速度和磁盘速度之间的差距，甚至超过了改进的处理器技术所造成的差距。本项目研究在这种大规模共享内存多处理器中管理不同级别共享存储的操作系统机制和策略。具体项目分为以下几类：主存管理(包括传统的虚拟内存和不同主存模块之间的页面迁移和复制)、并行文件系统设计、工作负载特征和性能评估方法以及编程环境在解决NUMA问题中的作用。

Nonuniformity of memory access is an almost inevitable feature of the memory architecture in shared memory designs that can scale to large numbers of processors, creating the increasingly important class of NUMA (nonuniform memory access time) multiprocessors. Managing the placement and movement of code and data is crucial to performance in NUMA architectures. This has been called the NUMA Problem. The I/O bottleneck problem is also exacerbated by parallel architectures which widen the gap between processing speed and disk speed even beyond that created by improved processor technology. This project investigates operating system mechanisms and policies for the management of the various levels of shared storage in such large- scale shared memory multiprocessors. The specific projects divide into the following classes: main memory management (both traditional virtual memory and the migration and replication of pages among various main memory modules), parallel file system design, workload characterization and performance evaluation methodology and the role of the programming environment in solving the NUMA Problem.