并行文件系统中适度贪婪的Cache预取一体化算法

PROPERLY GREEDY CACHE PREFETCH INTEGRATED ALGORITHM IN THE PARALLEL FILE SYSTEM

传统文件系统中的Ｃａｃｈｅ和预取技术是两种降低访问延迟的有效方法．在并行科学计算应用的Ｉ／Ｏ访问模式下，简单的Ｃａｃｈｅ和预取技术已无法提供较高的Ｃａｃｈｅ 命中率．该文在分析该Ｉ／Ｏ模式的基础上提出了适度贪婪的Ｃａｃｈｅ 和预取一体化算法（ＰＧＩ）．该算法充分利用了并行文件系统环境的特点，采用了适度贪婪的动态滑窗技术，可以有效地消除预取时的抖动，降低系统处理开销；并同时采用了Ｃａｃｈｅ 和预取一体化的淘汰损失估计算法，使淘汰的损失降到最低点，在整体上提供了较短的Ｉ／Ｏ

Cache and Prefetch is two efficient ways to lower the delay of I/O request. The latency of physical I/O operations can be avoided by caching frequently referenced data in primary memory rather than in secondary memory. And latency tolerance can be achieved by prefetch's overlapping long I/O access with independent operations. But simple cache and prefetch algorithm is no use for parallel scientific application, and prefetching must be carefully balanced against caching. This paper proposes a new prefetch algorithm faced to parallel scientific application: Properly Greedy Cache Prefetch Integrated Algorithm (PGI). PGI is based on the regularity of parallel scientific application's I/O access mode. Now, LRU _SP and TIP algorithm is too greedy when they execute prefetch operation. PGI can overcome the shortcomings of LRU _SP and TIP. PGI uses properly greedy dynamic slide window, which can adjust its prefetch size according to application's consume speed and I/O service delay. When replacing the cache blocks, PGI uses cache and prefetch integrated loss estimate algorithm to achieve the least loss and least I/O service time. In parallel file system environment, the workload of each I/O node is different, PGI thinks of it fully. When replacing cache blocks, PGI will first replace those cache blocks on light workload I/O nodes. So, the workload of each I/O node will be balance, and the total service time will be lower. The result shows that the PGI can bring higher hit ration, shorter I/O operation delay and balance workload.