Fast Operating System

Operating system (OS) emulation consists of running a binary executable intended for a particular OS on a different OS. This ability allows users to run more programs and OS manufacturers to jumpstart new OSs with non-native programs. There are several OSs that can run binaries compiled for the same processor but for different operating systems. There are also systems that can run multiple OS servers at the same time (e.g. Mach/microkernel architectures). Unfortunately, on these systems, one must either place the emulation code directly in the OS to gain speed or place the code in user space for flexibility and lose direct, and thus quick, access to OS internals. This thesis describes how one can take advantage of the exokernel architecture to make OS emulation both flexible, in the sense that it runs in user-space, and quick, in the sense that it can run fast. This combination has not been achieved on monolithic or microkernel architectures. To demonstrate the approach, this thesis describes a prototype emulator (XPOSE) that runs on an Intel-based exokernel (XOK) with a UNIX-like library OS (ExOS) and can run many programs compiled and linked for the monolithic OpenBSD UNIX OS. All emulation code is in user space and, because the emulator is linked to the library OS, it can access OS internals. Measurements show that XPOSE runs OpenBSD programs with very little overhead. Finally, the mechanisms XPOSE uses to intercept system calls makes it useful for debugging and perfomance analysis. Statistics gathered with XPOSE indicate where effort should be focused to reduce the number of kernel calls per library OS call. Sapuntzakis for helping me understand WINE. Also, I would like to thank DARPA for helping to fund this project. I dedicate this thesis to my parental units.