Visualizing Predictability of File Access Patterns

A. Luo, A. Amer, N. Der, D.D.E. Long, and A. Pang (USA)


information visualization, conditional entropy, caching ef fects


We propose a novel method to study storage system pre dictability based on the visualization of file successor en tropy, a form of conditional entropy drawn from a file ac cess trace. First-order conditional entropy is used as a mea sure of predictability. It is superior to the more common measures such as independent likelihood of data access. For file access data, we developed a visualization tool, VIP (Visualizing I/O Predictability) system, that produces 3D graphical views of the variation in predictability of suc cessive access events on a per-file basis. Our visualiza tion tool provides interactive observation of the variations in predictability according to some arbitrary criterion, e.g. time of day, cache scale, program identifier, user groups, or any other classification of files. Four entropy data sets were extracted from various file system traces. These four data sets are representative of the variability in file access patterns for different machine use: server, personal work station, large number of interactive users, and heavy write activity. Visualization results show that there is strong pre dictability among files and optimizations would be prof itable.

Important Links:

Go Back