Increasing Timing Resolution for Processes and Threads in Linux

The goal of the project was to develop a method for obtaining timing information for Linux processes and threads that is more precise than the one currently provided by getrusage(). The getrusage() system call is a fairly standard one, that returns the time a process has been running (CPU time) in terms of seconds and microseconds. With even the most recent builds of the stable Linux 2.2.14 kernel, the timing is obtained by sampling the internal clock tick of the Linux operating system. This type of sampling can lead to imprecision in the timing of processes and threads as the basic clock tick unit in Linux (the constant HZ) is only operating at 100Hz. With this 100Hz system tick, it is possible to only approximate times within 10ms of the actual time for each sampled time period. Thus several of the digits in the microsecond portion of the getrusage() return values are inaccurate from the beginning. The current implementation of getrusage() uses these ticks to calculate the actual user time spent on a process. Whenever data is collected in this fashion, there is always the possibility that a sampling error is induced. When the Linux kernel samples this 100Hz system tick, it is possible to obtain inaccurate timing results, as the process may swap in right before the tick and swap out right after the tick. This situation will then cause an incorrect 10ms to be added to the user time of the process, when in fact the actual time spent is much less than 10ms. With any long term process it is apparent that this imprecision in data acquisition might not amount to much of a difference, however with very fast operations, this error can add up quickly, and the accuracy of the timing can be affected greatly.