【正文】
gregated in a gridstyle environment. They are distributed across diferent institutions and connected on widearea works, providing the puting power and storage capacity to support executions of many applications. This dissertation focuses on two specific aspects of data management in distributed systems: data provisioning — providing applications running on the puting resources with remote access to their data stored on the storage resources, and the management of the data provisioning — the establishment, configuration, and termination of the remote data access. Computing in a gridstyle environment poses unique challenges to these tasks because of the above mentioned heterogeneous, dynamic, and largescale nature of applications and resources. First, the diversity of applications and resources motivates a data provisioning solution that can be transparently deployed, without modifying the existing operating systems (O/Ss) and changing the application source code or binaries. Second, the widearea, crossdomain environments necessitate applicationtailored optimizations for data access to address the inefciency (long work delay, limited work bandwidth), insecurity (insecure resources, limited mutualtrust between diferent domains), and unsafety (unreliable machines and works) that are typical in such environments. Last but not least, the management of data provisioning in a large, dynamic system also desires ?exible control and automatic optimization of the remote data access, in order to deal with the plexity of providing data to many applications, to agilely adapt to the changing environments, and to deliver applicationdesired performance, security, and reliability. To address these challenges, this dissertation presents a twolevel data management system in which file system virtualization provides applicationtailored gridwide data access, and servicebased middleware enables autonomic management of the data provisioning. In particular, this system has made the following contributions: It provides ondemand, crossdomain data access transparently for unmodified applications and O/Ss based on userlevel virtualization of widely available O/Slevel distributed file systems (DFSs). It supports applicationtailored enhancements designed for gridstyle environments on several important aspects of remote data access, including performance, consistency, security, and reliability. It employs middleware services to achieve ?exible and interoperable management of gridscale data provisioning, which is capable of controlling the lifecycles and configurations of dynamic data sessions based on application needs. It develops autonomic functions to automatically optimize the data management according to highlevel objectives, in order to reduce the plexity of managing data sessions and adapt them promptly to changing environments. Finally, the proposed system has been demonstrated, with thorough experimental evaluation, that it is efective and can significantly outperform conventional DFSbased approaches in gridstyle environments。s behaviors and needs. Because an optimization tailored for one application (., aggressive prefetching of fil