Ph.D. Dissertation. Carnegie Mellon University Parallel Data Lab Technical Report CMU-PDL-09-108, August 2009.
Andrew J. Klosterman
Department of Electrical and Computer Engineering
Carnegie Mellon University
Pittsburgh, PA 15213
The basic distributed, object-based storage system model lacks features for storage management. This work presents and analyzes a strategy for using existing facilities to implement atomic operations on sets of objects. These bulk operations form the basis for managing snapshots (read-only copies) and forks (read-write copies) of portions of the storage system. Specifically, we propose to leverage the access control capabilities, and annotations at the metadata server, to allow for selective clone and delete operations on sets of objects.
In order to act upon a set of objects, a bulk operation follows these steps. First, the metadata server accepts the operation, contacts the storage nodes to revoke outstanding capabilities on the set of objects, and retains a record of the operation and the affected set of objects. At this point, clients can make no changes to existing objects since any capabilities they hold will be rejected by storage nodes. Second, when clients subsequently contact the metadata server to access affected objects (e.g., acquire fresh capabilities), any records of bulk operations are consulted. Finding that a client is accessing an affected object, the metadata server will take the necessary steps to enact the uninstantiated operation before responding to the client request. This eventual enforcement of operation semantics ensures compliance with the operation’s intent but delays the corresponding work until the next client access. With appropriate background instantiation, the work of instantiating bulk operations can be hidden from clients.
In this dissertation, we present algorithms suitable for performing bulk operations over distributed objects using m-of-n encodings. The core logic is concentrated at the metadata server, with minimal support at clients and storage nodes. We quantify the overheads associated with the implementation and describe schemes for mitigating them. We demonstrate the use of bulk operations to create snapshots in an NFS server running atop distributed, object-based storage.
FULL Thesis: pdf