Maintaining bounded disorder files in multiprocessor multi-disk environments

Abstract

The Bounded Disorder file is a file organization method that consists of an index and multi-bucket data leaves which are partially expanded before being split. It guarantees good performance for single-key operations (e.g. search, insert, and delete), similar to that of hash based methods, and in addition, good performance for subrange operations, and reasonable performance for sequential-key operations. To obtain these capabilities, the index must be entirely stored in main memory.

In this paper, we consider the problem of implementing Bounded Disorder files in multiprocessor multi-disk environments which consist of P processor-disk pairs. The processors are either tightly coupled (communicate via shared memory) or loosely coupled (communicate via a local network).

The straightforward solution is to equally partition the file records among processors, each of which maintains its part as a “local” Bounded Disorder file (stored in the processor's main memory and disk). This method is highly parallel (up to P single-key operations can be executed in parallel) and achieves good performance due to the use of Bounded Disorder files.

We present an alternative method, called Conceptual Bounded Disorder file, which obtains performance similar to the above straightforward method, and in addition, obtains a significant cut down in main memory space consumption. Thus, in on-line systems, the number of files that can be concurrently opened for access is much larger when files are maintained as Conceptual Bounded Disorder files.

This research was supported by the E. & J. Bishop Fund at the Technion.