Declustering of key-based partitioned signature files

Reviewer: Jaroslav Pokorny

Declustering data on multiple disk systems has become an important research issue concerning partial match and range queries on both formatted and unformatted data. The paper proposes two parallel signature file organizations, Hamming filter and Hamming + filter. These organizations are based on parallelization of the quick filter (QF) key-based organization. Parallelization of QF can be considered as the problem of declustering buckets on disks so that, for each possible query, the I/O load is distributed uniformly among the available disks. In section 2, the authors give basic concepts related to binary linear error-correcting codes, key-based partitioned signature files, and Hamming filter (HF) organization [1]. This dynamic key-based signature file organization for multiple disks combines the dynamic clustering features of QF with a static declustering strategy based on linear error-correcting codes. Section 3 is devoted to an analysis of HF performance. Optimal, expected, and worst-case response times are derived. Further, it is shown that the performance of HF depends on the file size. Section 4 defines Hamming + filter (H + F), a new organization that guarantees high performance by automatically adjusting the static disk allocation strategy used in HF. The basic idea is to use variable-length keys not only to manage the signatures stored on each disk, but to assign signatures to disks. This requires using a sequence of error-correcting codes and one code for specific file size; it results in a smooth migration of signatures between disks, so no reorganization phase is needed. Section 5 summarizes experimental results on HF and H + F obtained by prototype systems developed by the authors. The tests focus on the quality of declustering, that is, how evenly the organization can distribute the I/O load on disks. H + F can guarantee good performance levels for even highly dynamic files, justifying the migration process that linearly reallocates signatures when the file grows or shrinks. Section 6 discusses related work and states some conclusions. The paper concludes with a number of appendices that contain the proofs of the most important theorems. The paper is written at a high level and significantly contributes to the theory of signatures. The results are also of practical importance.