ACM International Conference on Supercomputing (ICS) 2025

Paper Accepted at an Outstanding International Conference

(Safdar Jamil, Ph.D., post-doctoral researcher)

The paper 'DEDUPKV: A Space-Efficient and High-Performance Key-Value Store via Fine-Grained Deduplication', written by Safdar Jamil, Ph.D., post-doctoral researcher at the Data-Centric Computing and AI Systems Laboratory (DISCOS) (first author), and Professor Kim Young-jae (corresponding author), has been accepted for publication at the ACM International Conference on Supercomputing (ICS 2025).

Meta and ByteDance have widely adopted RocksDB as the de facto storage engine for applications requiring high throughput in cloud environments. Most of the key-value data they use, however, shows a high rate of duplication. Such duplicate data greatly reduces storage efficiency. Moreover, because key-value data is typically very small — on the order of tens of bytes, far smaller than the 512-byte to 4KB access units of traditional block or file systems — file-system-level deduplication has not been effective.

To resolve this problem, Dr Safdar Jamil (first author) and Professor Kim Young-jae (corresponding author) carried out the world's first research on fine-grained, value-level deduplication in RocksDB, and presented the results at ICS 2025 (the ACM International Conference on Supercomputing). Since there had been no prior work at all on data deduplication in RocksDB, this research is highly significant as the first attempt in the field.

In the paper the authors propose DEDUPKV, a deduplication-capable LSM-tree based key-value store, and implement it in BlobDB, which applies RocksDB's key-value separation technique. By eliminating duplicate writes, DEDUPKV improves both throughput and storage efficiency: up to 33% higher throughput than BlobDB (the baseline) on write-intensive workloads, more than a twofold reduction in write amplification, and up to 45% storage savings on real datasets. The detailed design of DEDUPKV is shown in the figure below.

< Design overview of DEDUPKV>

Professor Kim Young-jae, Dr Safdar Jamil's supervisor, said: "This paper is the fruit of research Dr Safdar has quietly and steadily pursued throughout his doctoral course, and its acceptance at a globally prestigious conference is deeply meaningful and gratifying," adding, "This achievement is the result of his determination to keep attempting new observations as a researcher, and I sincerely congratulate the first author, Dr Safdar Jamil."

ACM ICS is recognized as one of the world's most prestigious conferences in high-performance computing and systems, and is listed among the outstanding conferences in the software field of the Korean Institute of Information Scientists and Engineers.

References:

·         ACM International Conference on Supercomputing (ICS) 2025

Website: https://hpcrl.github.io/ICS2025-webpage/home.html