山东大学-proj45-RLWM-3712

Project information

相变存储器（PCM）具有存储密度高、静态漏电功耗低以及接近 DRAM 的访问延迟等优势，但在写入过程中，尤其是执行 RESET 操作时，由于热扩散效应，仍然容易产生写扰动，对相邻存储单元造成干扰。现有的写扰动缓解方法大多依赖于纠错机制、编码技术或固定策略，因此难以根据工作负载和数据分布的变化进行动态调整。针对这一问题，本项目提出了一种基于强化学习的异地更新（Out-of-Place Update）方法 RLWM。RLWM 将写地址选择建模为一个强化学习问题，由智能体根据当前写入模式和失效数据块（stale block）的分布，选择能够最小化写扰动的目标位置。通过轻量级在线训练，RLWM 能够根据工作负载变化不断调整策略，实现长期、持续的写扰动抑制。实验结果表明，RLWM 平均可降低 46% 的写扰动，相较于SOTA进一步降低了 16%。这说明RLWM 能够在较低额外开销下，有效缓解 PCM 的写扰动问题。

Phase Change Memory (PCM) offers high density, low leakage power, and near-DRAM latency, but remains vulnerable to write disturbance, especially during RESET operations where thermal diffusion may disturb adjacent cells. Existing mitigation methods rely on error correction, coding, or fixed heuristics, and therefore adapt poorly to changing workloads and data patterns. We propose RLWM, a reinforcement-learning-based out-of-place update scheme that mitigates write disturbance in PCM. RLWM formulates write-address selection as a reinforcement learning problem, where the agent selects a target location that minimizes write disturbance based on the current write pattern and stale-block state. Through lightweight training, RLWM adapts online to workload changes and achieves long-term write-disturbance mitigation. Experiments show that RLWM reduces write disturbance by 46% on average and achieves a further 16% reduction compared with the state-of-the-art method. These results demonstrate that RLWM provides effective write-disturbance mitigation for PCM with low additional overhead.