We described a technique called bundling which splits a collection of class files and resources into bundles based on previously observed class and resource loading behavior. We showed that bundling is competitive with cumulative compression when the applications and profiles are known in advance, and that it is no worse than the Jar archive format when used on an application not included in the training set.
One possibility for future work is to use a static dictionary of commonly occurring class file data in compressing the bundles. This could improve the compression ratio on small bundles.
Another possibility for future work is to apply bundling in other contexts where the transfer of sequences of files is required. For example, it might be applicable in serving a collection of files to be partially mirrored at other sites.