| Summary: | 碩士 === 國立交通大學 === 電子研究所 === 107 === As big data analysis and deep learning have well developed, storage system is also more important. Because data that we need to deal with are bigger and bigger, efficient disk usage and faster read/write rate is significant issue for user. In recent years, learning methods have been applied in many software issues. However, learning methods on hardware issues are lightly explored.
We take advantage of deep learning algorithm to improve efficiency and speed of solid-state disk. We focus on two issues. First, we divide address of data into hot type and cold type. For different types, we can adjust priority of usage of disk to achieve optimal utility. Applicable allocation for hot/cold address increases the lifetime of SSD and better write throughput. For second issue, we predict next delta address and construct accurate and efficient address prefetchers. It is directly helpful to reduce data access time and increase read throughput.
For the former topic, we achieve about 93.5\% accuracy by time-dependent NN model. The lifetime increase 6930 cycles and write throughput increase 39.43MB/s. For the latter topic, we construct three-channels network and simulate cache mechanism to predict next delta address. In this topic, read throughput increases 156.98MB/s by our model. Our work represents learning methods work well on hardware architecture. In the experiment part, simulation results show that our proposals work better than traditional SSD.
|
|---|
