The ideal approach would be to run identical applications on both Linux and Windows NT platforms. Unfortunately, we do not have such applications for a variety of reasons. First, our OpenGL applications are written for the NT platform because it is the single most popular platform that all high-end graphics accelerator vendors support. Second, the shared-memory applications currently use an SVM library that was solely developed for the Win32 platform, primarily due to its mature support for kernel threads. The incompatibility between the Linux and Win32 API also made it very difficult to have a single SVM protocol library that works for both OSes.
Since we do not have the identical applications on the two platforms, we perform the evaluation as follows: First, we use microbenchmarks to measure the various aspects of the communication system and the relevant aspects of operating systems on the two platforms. Then we instrument the applications to measure the frequency of occurrence of the various events like UTLB misses and notifications on applications running on windows NT platform. Finally, we combine the two measurements to predict the impact of the OS on the applications performance.