File, Database and I/O Systems Performance

有两种技术趋势强烈地影响I/O系统的设计和性能。首先，处理器的能力，特别是对于单（或少数）芯片设计，正在迅速增加，导致：(a) I/O负载（几乎）相应的增加，(b)具有工作站（没有两个本地磁盘）、文件服务器和计算节点的分布式系统，以及(c)与单处理器相比，多处理器具有巨大的成本/性能优势，提供相同的总性能。其次，磁盘存储密度正在迅速增加，而磁盘性能却没有相应的提高。由于这些趋势，一些设计和优化问题变得极其重要：(A)系统越来越容易受到输入/输出瓶颈的影响，因为输入/输出速率不断增加，而输入/输出系统对这些瓶颈的反应能力的增长速度却越来越慢。(b)具有共享I/O设备和/或数据库的多处理器系统在引用、锁定和更新共享数据方面越来越容易受到I/O瓶颈的影响。该项目继续研究与文件、I/O和数据库系统性能有关的许多主题。这些包括：(a)开发和分析分布式和分层系统中磁盘缓存的算法，同时考虑到一致性和性能。(b)研究磁盘缓存结合磁盘条带化和日志结构文件系统的影响。(c)分析和发展共享的分布式数据库系统的锁定算法。(d)将在加州大学伯克利分校建造的使用光盘的超大容量存储系统中分层和分布式文件迁移的设计、优化和实现。这项工作很大程度上依赖于实际系统测量和轨迹驱动仿真。

There are two technology trends strongly impacting the design and performance of I/O systems. First, processor power, especially for single (or few) chip designs, is increasing very rapidly, leading to: (a) (almost) corresponding increases in I/O load, (b) distributed systems with workstations (w/wo local disks), file servers and computation nodes, and (c) multiprocessors with enormous cost/performance advantages over uniprocessors providing the same aggregate performance. Second, disk storage densities are increasing rapidly without any comparable increase in disk performance. A number of design and optimization problems have become extremely important due to these trends: (a) Systems are becoming increasingly susceptible to I/O bottlenecks due to increasing I/O rates and the less quickly in creasing ability of the I/O system to respond to them. (b) Multiprocessor systems with shared I/O devices and/or databases are increasingly susceptible to I/O bottlenecks in referencing, locking and updating shared data. This project continues research on a number of topics relating to file, I/O and database system performance. These include: (a) Development and analysis of algorithms for disk caching in distributed and hierarchical systems, taking into account both consistency and performance. (b) Studies of the effect of disk caching in conjunction with disk striping and log structured file systems. (c) Analysis and development of locking algorithms for shared, distributed database systems. (d) Design, optimization and implementation of hierarchical and distributed file migration in a very large mass storage systems using optical disks to be built at UC Berkeley. The work relies heavily on actual system measurement and trace driven simulation.