With all blocking eliminated and with a much higher request rate, we return to the issue of CPU consumption. By storing the CallTime field of each system call, we can track per-call performance by invocation, both to observe trends and to identify time-related problems. Traditional profiling tools usually report average CPU consumption of each function, thus hiding any performance trends. User space timing functions may catch the general trend in spite of the measurement error, but involve much more work to track each system call and find the problematic ones.