High-Performance File Systems for Scientific Multiprocessing

在这个项目中，开发和实现了一个可扩展和可移植的文件系统框架，使高级I/O操作成为可能。该项目使用各种并行架构的分析模型和模拟来评估策略和机制。当今并行计算机上的大规模并行计算受到其文件系统的限制。通过巧妙的硬件和软件支持，可以提高文件和磁盘系统的性能。通过使程序员能够在高层指定I/O请求，可以提高I/O性能。这类似于许多并行语言中的矩阵操作如何向编译器指定高级操作所需的语义，这样编译器、运行时系统、操作系统和硬件就可以协调I/O，以优化访问磁盘的数量和顺序、缓冲区的使用、磁盘上文件的布局等等。这些在更高级别指定的技术可以推广到其他形式的I/O，比如网络I/O。

In this project, an extensible and portable framework for a file system to make high-level I/O operations feasible is developed and implemented. The project evaluates the policies and mechanism using both analytical models and simulations of a variety of parallel architectures. Massively parallel computations on today's parallel computers are limited by their file systems. The performance of the file and disk systems can be improved by clever hardware and software support. I/O performance can be improved by making it possible for the programmer to specify I/O request at a high level. This is similar to how the matrix operations in many parallel languages specify the required semantics of high level operations to compilers, such that the compilers, run-time systems, operating system and the hardware coordinate the I/O to optimize the number and the order of accesses to disks, usage of buffers, layout of files on disks, and so on. These techniques of specifying at higher level can be generalized to other forms of I/O like the network I/O.