DStream: Dynamic Memory Resizing for Multi-Streamed SSDs

To reduce the write amplification factor (WAF) and to improve the performance of flash memory SSDs, the multi-stream SSD (MS-SSD) has been recently proposed, which writes different streams into separate flash memory blocks. The previous studies on MS-SSDs focused on only how to divide user requests into different streams assuming that MS-SSDs support a fixed number of streams. In this paper, we introduce a dynamic stream number adaptation technique, called DStream, which can improve the performance of MS-SSDs. By adjusting the number of streams adaptively according to the workload with k-means clustering, the remaining memory resources can be used for the map table cache to improve SSD performance. Experiments based on MS-SSD simulator show that DStream reduces cache miss rates by up to 82% and the number of flash memory I/Os by 24%.