4.  Conclusions

      We have created a profiler that aids in the evaluation of the kernel. For each routine in the kernel, the profile shows the extent to which that routine helps support various abstractions, and how that routine uses other abstractions. The profile assesses the cost of routines at all levels of the kernel decomposition. The profiler is easily used, and can be compiled into the kernel. It adds only five to thirty percent execution overhead to the kernel being profiled, produces no additional output while the kernel is running and allows the kernel to be measured in its real environment. Kernel profiles can be used to identify bottlenecks in performance. We have shown how to improve performance by caching recently calculated name translations. The combined caches added to the name translation process reduce the average cost of translating a pathname to an inode by 35%. These changes reduce the percentage of time spent running in the system by nearly 9%.