Estimating File Compressibility Using File Extensions

Abstract

Adaptive compression systems dynamically choose a compression strategy — including no compression — by monitoring CPU usage, output rate, expected time to compress, and perhaps most importantly, the estimated compressibility of the data. Many adaptive compression systems were designed with the assumption that files with the same filename extension will compress roughly to the mean compression ratio (the ratio of compressed size to original size) of some set of files with the same extension. This implies that the compression ratio distribution follows a normal distribution. Though a normal distribution of compression ratios may seem intuitive, this assumption lacks strong empirical supporting evidence. To test this assumption, we built a tool to compress real-world files from many participants, storing the compressed size, original size, file extension, and other metadata. The results of three tests for normality indicate that none of the file extensions we analyzed have a normal distribution, though for some extensions, not all three tests agree. Furthermore, quantitative analysis reveals that files with the same extension compress according to multiple different distributions, and we identified some readily accessible metadata that can separate these files into simpler distributions. We conclude with a discussion of the utility of mean compressibility as an estimator and the implications this study has for future research in adaptive compression.