嵌入式系统IO存储栈强化学习最优控制

作为嵌入式系统最重要的子系统之一，IO存储栈的设计直接影响系统性能与闪存寿命。当前IO存储栈大量使用静态IO管理技术，在复杂的IO存储环境中会导致性能次优与闪存寿命缩短。本项目探索基于强化学习来解决复杂环境下的自适应最优IO决策问题，主要包括以下三个内容：1）繁重的写压力会严重影响嵌入式系统IO实时性能与闪存寿命，本项目提出学习IO特征来得出IO负载自适应的最优压缩算法；2）NVMe闪存设备多队列机制在嵌入式系统中存在队列资源分配不合理与缺乏QoS保证机制的问题，本项目提出学习IO特征和系统实时状态来得出队列资源分配与IO实时调度的最优决策；3）多流闪存设备存在IO数据生命周期识别精确度与流资源分配机制的问题，本项目提出学习生命周期相关IO特征与流资源分配相关环境变量来实现IO流精确归类与流资源最优分配决策。综合以上技术，本项目可以实现系统性能与寿命均衡最优。

With the emerging technology of IoT (Internet-of-Things), embedded system has been one of the most attractive research areas. As an important sub-system in embedded devices, the design of I/O stack could dominate the system performance and flash storage lifetime, thus impacting the user experience and cost of the embedded devices. Bounded by the development of the traditional computer technology, static managing schemes are adopted in the current I/O stack, which could not adapt to the complicated environment in the I/O stack sub-system, thus incurring sub-optimal system I/O performance and shortened storage lifetime. To mitigate this issue, this project proposes reinforcement learning-assisted I/O managing approaches, including the following contents: 1) Excessive write I/O could directly impact the system I/O performance and shorten the lifetime of flash storage devices, this project proposes to reach the self-adaptive optimal compression algorithm through learning related I/O patterns; 2) The current multi-queue I/O scheduling mechanism in embedded devices could incur the sub-optimal queue resource allocating efficacy meanwhile lack of QoS guarantee solution, this project proposes to acquire the optimal decision in queue resource allocation and I/O scheduling through learning related I/O patterns and the real-time status of the I/O system; 3)The I/O data lifetime clustering and flash resource allocation mechanisms in the current multi-stream flash device are problematic, this project proposes a learning-based I/O data lifetime clustering solution to improve the clustering accuracy, and a reinforcement-learning based flash resource allocating approach to reach the optimal flash resource allocating decision. Combining with all these techniques, this project could reach the Pareto-optimal system performance and flash device lifetime.