Understanding Filesystem Performance for Data Mining Applications
Bouchra Bouqata, Christopher D. Carothers, Boleslaw Karol Szymanski, Mohammed Javeed Zaki · 2007
Motivated by the importance of I/O performance in data mining efficiency, we focus this paper on analyzing data mining performance across different file systems. In our study, we consider three of the most popular filesystems available under the Linux distribution. These include: Ext2[3] (nonjournaled), Ext3[16] (journaled), and Reiser[12] (journaled). We conclude that full data and metadata journaling (Ext3) appears to dramatically slow down the performance of a data mining engine. For I/O intensive cases, data mining execution time was double when run over Ext3 as compared to Reiser. We found that the write speed of Ext3 was 35 times slower than Reiser, and file address references display only a shortrange dependence in data mining, indicating high degree of locality in data references.