System Support for High Performance Computing

点击翻译按钮获取中文摘要

The fundamental issue addressed in this work is the structuring of operating systems to take advantage of multiprocessors so as to yield the best possible application performance. The division of labor between the kernel and the user level is different on a multiprocessor than on a uniprocessor. On a uniprocessor, the kernel is responsible for address spaces, communication, resource allocation, threads, and thread scheduling. On a shared memory multiprocessor, the kernel iis left in charge of address spaces and resource allocation, but communication, threads, and thread scheduling are implemented at user level through runtime support in each address space. Furthermore, each address space is in direct control over certain key aspects of virtual memory management and file management. It is important to understand that while ``micro-kernel'' designs such as Mach 3.0 take the important step of moving the file system, the virtual memory system, etc., out of the kernel to user-level servers, they leave cross-address space communication, threads, and thread scheduling in the kernel. There are structural differences between these monolithic kernels, micro-kernel designs and the ``nano-kernel'' advocated here. In experimental work over the past several years, the researchers have constructed proof-of-concept prototypes in most of the essential areas that must be addressed: locking protocols, user-level threads, cross-address space communication, the kernel interface, processor allocation policies, and user-level virtual memory management. The objectives in this work are of two kinds, direct and indirect. Direct objectives include: (1) integrating these proof-of-concept prototypes into a single system, convincingly demonstrating the viability of the overall approach; and (2) solving the problems that stand in the way of this integration, some of which are known and some of which can only be discovered by undertaking the task. Indirect objectives include: (a) using the ``core'' of the resulting system to study a number of important policy questions (thread scheduling, processor allocation, virtual memory management) that can only be explored given such an experimental testbed, and (b) using this core as a base for exploring high performance system support for more advanced architectures, specifically large scale machines with distributed memory, distributed/parallel programming using high band-width low-latency networks, and wide-address machines.