Deep contrastive learning–based redundancy elimination for similar data in cloud storage systems

Cloud storage systems contain a large amount of both duplicate and similar data, which imposes significant storage overhead. Redundancy elimination has become an important technique for improving storage efficiency. However, existing redundancy elimination methods still lack deep semantic understanding and expressive feature representations, especially for identifying similar blocks beyond exact duplicates. To address these limitations, this paper proposes DCLC, a redundancy elimination method based on deep contrastive learning. DCLC constructs a semantic representation model that integrates local features with global contextual information through self-supervised contrastive learning. Therefore, it generates more discriminative embedding vectors and improves the accuracy and efficiency of identifying similar blocks. Furthermore, a dynamic cache mechanism based on access locality processes most queries for similar blocks directly in the cache, which avoids unnecessary vectorization and approximate nearest-neighbor searches. Finally, DCLC arranges multiple candidate base blocks in sequence to form a cooperative reference unit, enabling the delta encoder to exploit long-range structural correlations and eliminate cross-block redundancy. Experimental results demonstrate that, compared with representative existing approaches, DCLC improves system throughput by 4.18 \(\times \) to 59.58 \(\times \) and achieves a maximum improvement of 248% in overall compression ratio under the evaluated workloads. The source code is available at: https://github.com/qcode-systems-lab/DCLC .