A Proposal for a New Data Model and its Implications on I/O Architecture

摘要

Today, applications are being developed with little regard to the needs and capabilities of I/O. The root cause of poor performance is not due to I/O but rather virtual disregard of I/O by using applications. By adjusting how we view, store, and process data so as to be more in consonance with I/O devices, significant performance gains can be achieved. This paper examines the performance issues of non-numeric applications that are primarily I/O bound. By preserving the logical topology of data when mapping to physical disk, we can exploit the temporal and spatial properties of data being systematically processed. Unlike the traditional data model, data access tends to be more granular, synchronous and more efficient. Using intelligent controllers, each with a set of spindles, a high degree of parallelism can be achieved without the overhead usually associated with parallel architectures.