Clustering data according to update frequency to reduce garbage-collection overhead in solid-state drives

Garbage collection, which entails multiple page copies and a block erase, is a major source of performance fluctuation and degradation for NAND flash memory-based solid-state drives. This work aims to reduce its overhead by generating a skewed distribution of valid pages over all the blocks. Therefore, we propose classifying data as hot, warm, or cold according to their update frequencies, and to cluster them into different blocks. Our performance evaluation shows that the proposed scheme reduces the total garbage-collection count up to 43.4%, compared to the original page-mapping scheme, for an average performance improvement of up to 34.5% without any additional memory overhead.