August 1st, 2002
Categories: Networking
With the large diversity of network resources, computing resources and heterogeneous devices, there is a problem when there is a mismatch between network-capacity and the size of data transferred, or between the type of data a device can handle and the type of data transferred. Traditional distributed file systems (DFS) including NFS [1], Andrew File System (AFS) [2] and Coda [3], fail to address this issue primarily because they cannot adapt their services according to environmental constraints. The fundamental problem is that data (as files) is a stream of bytes carrying no information by itself, requiring external applications to be meaningful, making it impossible for systems to make decisions effective for the data. Earlier approaches with semantic file systems focused only on extracting information from the data for inference engines. Object oriented operating systems, like Choices [4] and Apertos [5], do not deal with the problem of system adaptation.
We address the problem of adaptation in a DFS by modeling the file system using persistent distributed objects. We propose an object-oriented framework, where files (persistent objects) in the DFS participate when the system reacts to environmental or user-imposed constraints. Our prototype simulation allows the system to adapt to environmental constraints like network congestion. When a video clip is played on a client, the DFS transparently chooses a version with fidelity appropriate to the available bandwidth. If the network is congested, the user sees a low-resolution or low-frame-rate video. The system can also adapt to user-imposed constraints like security and distribution-rights. For the same example, if the network is untrusted, the client can obtain an encrypted version of the clip (currently, the system cannot do this transparently). If there are copyright restrictions, the system controls the distribution of the video to clients.
Our results show that the response time of the file system can be arbitrarily improved (at the expense of data-fidelity), which is impossible in NFS or AFS. Our design allows for implementing specialized security policies at a fine granularity (groups of files). With a minimal set of application interfaces to conform to, the framework can be extended and customized effectively for specific application domains.
Kumar, A., Theys, M. D., Object-Oriented Framework for Adaptation in a DFS, Proceedings of SPIE - Java / Jini Technologies and High-Performance Pervasive Computing, pp. 99-108, August 1st, 2002.