Varmosa : just-in-time binary translation of operating system kernels

Abstract

This thesis presents a just-in-time binary translation scheme that dynamically switches between system emulation with a slower but more memory efficient instruction interpreter, and a faster, more memory intensive binary translator. In testing, this hybrid interpreter/translator scheme reduced the size of the binary translation cache by up to 99% with a slowdown less than a factor of 5x in the worst case, and less than a 2x in the best case compared to a pure binary translation scheme. With only a 10% decrease in performance, upwards of 49% memory reduction is demonstrated. Additionally, a technique of guest kernel introspection and profiling using binary translation is presented.