An Adaptive Migration Method for Hierarchical Storage Data Based on Deep Reinforcement Learning

Abstract

The invention belongs to the field of big data storage, and relates to a hierarchical storage data adaptive migration method based on deep reinforcement learning. Based on the idea of ​​the deep reinforcement learning DQN model, combined with the characteristics of the hierarchical storage system, the state space, action space, and reward value are defined, and an adaptive data migration method is designed and implemented. Migration decision is made under the guidance of the system, and finally the system performs corresponding data migration according to the decision. The self-adaptive migration algorithm designed by the invention improves the throughput of the hierarchical storage system, provides lower delay, fully utilizes the advantages of SSD storage devices, reduces storage costs, and improves the data access performance of the hierarchical storage system.

Description

technical field [0001] The invention belongs to the field of big data storage, and relates to a hierarchical storage data adaptive migration method based on deep reinforcement learning. Background technique [0002] The advent of the big data era puts forward higher requirements for read and write performance on storage technology, and at the same time, storage devices are constantly being introduced. In addition to ordinary disk HDDs, solid-state disks (SSDs) and non-volatile memory NVMs have also appeared. The read and write delays of HDDs are at the millisecond (ms) level, and consume a lot of power, but the data storage has a relatively long durability and the storage price is very low, about 0.2 yuan / GB; while the read and write delays of SSDs are microseconds (μs) level, the read and write speed is relatively fast (especially the random read performance is high), but the price is relatively high, about 1 yuan / GB; and the Intel Optane series storage 905P, which has been...

AI Technical Summary

Problems solved by technology

However, since hierarchical storage systems often support a variety of different applications, different applications have very different data access modes, and the distribution often varies greatly. The page replacement algorithm of the one-size-fit-all mode often cannot cope with this situation well.

Images