存储系统中的频繁访问模式挖掘

Mining frequent access patterns in storage systems

研究、分析了影响经典的模式挖掘方法挖掘频繁访问模式的效率,使其难以被存储系统接受的主要因素——噪音的产生原因和表现类型,提出一种具有噪音过滤能力,适应存储系统频繁访问序列模式挖掘的新方法——Z-Miner。Z-Miner使用全局分支裁剪和分支聚类方法来过滤噪音,对实际系统工作负载的模拟结果显示,Z-Miner指导的预取可以使缓存失效率降低40%～66%,平均响应时间降低26%～66%。相对经典挖掘方法,Z-Miner的挖掘开销有1～2个数量级的下降,而预取优化效果提高了1倍。

Based on the analysis of the effect mechanism of the noise, a major factor that lowers the efficiency of frequent access pattern mining and makes classic mining methods unacceptable for storage systems, this paper proposes a novel mining method Z-Miner. The Z-Miner employs a global-branch-cutting and branch-clustering approach for noise filtering. The simulation results under real workloads show that the prefetching directed by the Z-Miner could reduce the cache miss ratio by 40 % - 66 %, and the average response time by 26 % - 66 %. Compared with classic mining methods, the overhead of the Z-Miner is 1 to 2 orders of magnitude less, while the efficiency of the prefetching is two times more.