First Name : Joseph
Last Name : Hellerstein
Title : Prof.
Company : UC Berkeley
Address : 387 Soda Hall #1776
City : Berkeley
State : CA
Mail Code : Telephone : 510-643-4011
FAX : 510-642-5615
E-mail : email@example.com
Today's information systems are based on designs that are a quarter-century old: centralized servers with local metadata, black-box user interactions, and underlying hardware assumptions frozen at the date of initial design. At Berkeley we are designing a new system called Telegraph. Our goal is a universal system for information: a global federation of information stores inter-operating seamlessly and autonomously, removing distinctions between databases, file systems, and the world-wide web. Telegraph consists of four composable components: a storage manager, an adaptive parallel/distributed query processor, an economic federator, and a progressive (CONTROL) interface combining queries, browsing, and scalable spreadsheets.
For HPTS we focus on the Telegraph storage manager, which is based on large variable-length segments. By borrowing from the past (Multics, the Burroughs B-5000), we (1) address the exponentially-growing bandwidth potential lost during seeks, and (2) allow for large sub-databases to be materialized and tuned for main-memory. We use a flexible interface to support a variety of recovery schemes, ranging from multiple-representation (a la write-ahead logging) to unified-representation (no distinction between "log" and "database".) Among unified schemes, we consider the possibilities of (1) leaving main-memory representations intact on disk ("database-only") and employing versions as needed, and (2) maintaining only a log on disk ("log-only") and "swizzling" pieces of log to and from main-memory format. Replication is built into the system; single-representation schemes does not imply a lack of physical redundancy.
The Telegraph storage manager will be used by the Ninja "internet services" OS/Networks project at Berkeley. The storage manager interfaces include a Linux file system, an HTTP server, and database iterators. Since our initial "customers" include filesystem applications (Ninja) and the query engine, we expect to tune for a wide range of workloads.
Joint work with Eric Brewer, Mike Carey, Steve Gribble, Mohana Krishna
Lakhamraju, and Mike Stonebraker.