Mitigating the Write Amplification Problem of Write-Optimized File Systems on Flash Storage

Abstract

As the volume of data stored by Big data and Cloud services continues to grow, both academia and industry are seeking for high-performance storage systems. Recently, with the recent advances in write-optimized indexes (WOI), WOI-based file systems can now outperform conventional file systems with orders of magnitude on random writes, metadata updates, and small file creation. Based on the B-tree structure, WOI-based file systems can not only process data faster than the conventional B-tree but also improve the range query performance. However, the write amplification of these WOI-based file systems becomes a serious performance overhead when adopting flash storage as underlying storage devices due to the recursive entry update behavior. To mitigate the write amplification problem of WOIbased file systems, we propose a flash-friendly WOI design to reduce the number of write requests on flash storage. To evaluate the performance of the proposed design, we adapt B+-tree as a case study and the experimental results are promising.