Efficiency and Scalability Issues in Metric Access Methods
| Authors | |
|---|---|
| Year of publication | 2008 |
| Type | Chapter of a book |
| MU Faculty or unit | |
| Citation | |
| Description | The metric space paradigm has recently received attention as an important model of similarity in the area of Bioinformatics. Numerous techniques have been proposed to solve similarity (range or nearest-neighbor) queries on collections of data from metric domains. Though important representatives are outlined, this chapter is not trying to substitute existing comprehensive surveys. The main objective is to explain and prove by experiments that similarity searching is typically an expensive process which does not easily scale to very large volumes of data, thus distributed architectures able to exploit parallelism must be employed. After a review of applications using the metric space approach in the field of Bioinformatics, the chapter provides an overview of methods used for creating index structures able to speedup retrieval. In the metric space approach, only pair-wise distances between objects are quantified, so they represent the level of dissimilarity. The key idea of index structures is to partition the data into subsets so that queries are evaluated without examining entire collections -- minimizing both the number of distance computations and the number of I/O accesses. These objectives are obtained by exploiting the property of metric spaces called the triangle inequality which states that if two objects are near a third object, they cannot be too distant to one another. Unfortunately, computational costs are still high and the linear scalability of single-computer implementations prevents from searching in large and ever growing data files efficiently. For these reasons, we describe very recent parallel and distributed similarity search techniques and study performance of their implementations. Specifically, Section 12.1 presents the metric space approach and its applications in the field of Bioinformatics. Section 12.2 describes some of the most popular centralized disk-based metric indexes. Consequently, Section 12.3 concentrates on parallel and distributed access methods which can deal with data collections that for practical purposes can be arbitrary large, which is typical for Bioinformatics workloads. An experimental evaluation of the presented distributed approaches on real-life data sets is presented in 12.4. The chapter concludes in Section 12.5. |
| Related projects: |