A High Effective Indexing and Retrieval Method Providing Block-Level Timely Recovery to Any Point-in-Time

Abstract

Block-level continuous data protection (CDP) logs every disk write operation so that the disk can be rolled back to any arbitrary point-in-time within a time window. For each update operation is time stamped and logged, the indexing for such huge amounts of records is an important and challenging problem. Unfortunately, the conventional indexing methods can not efficiently record large numbers of versions and support instant “time-travel” types of queries in CDP. In this paper, we present an effective indexing method providing timely recovery to any point-in-time in comprehensive versioning systems, called the Hierarchical Spatial-Temporal Indexing Method (HSTIM). The basic principle of HSTIM is to partition the time domain and the production storage LBAs into time slice and segments respectively according to update frequency of disk IOs, and build separate index file for each segment. In order to meet the demands of instant view of history data, the metadata of production storage is independently indexed. For long-time history data retrieval requirements, index snapshot is introduced in HSTIM to reduce the retrieval time. Another distinctive feature of HSTIM is its incremental retrieval method, which achieves high query performance at time point t + t if neighboring time point t is queried previously. The paper compares HSTIM with traditional B+-tree and multi-version B-tree (MVBT) index in many aspects. Experiments with real workload IO trace files show that HSTIM can locate history data within 8.05 seconds for recovery point of 48 hours, while B+-tree consumes 24.04 seconds. If the index snapshot is applied, HSTIM can reduce such retrieval time within 3 seconds.