The physical measurement of a system is useful up to a point - it can identify bottleneck devices and let you know when your system is becoming congested, but it can't tell you which users or applications are responsible. The UNIX 'ps' command will provide an accurate measure of the CPU usage of a process provided the process is a long-running one, and can be seen in several consecutive 'ps' snapshots. It will tell you nothing about short-lived processes (which can number in the hundreds of thousands in a day) falling in between successive 'ps' views.

UNIX accounting writes a record for each terminated process (about 40 bytes) and thus provides an unbroken record of all activity and resource consumption from the user/process point of view. It allows accurate measurement of not just the short-lived processes' resource consumption, but is the only means whereby the full story can be told of a long-running job. See below:

09:00     09:15     09:30     09:45     10:00     10:15
       |--------------------------------------|
  ^         ^         ^         ^         ^         ^  
  ps        ps        ps        ps        ps        ps
  no        yes       yes       yes       yes       no 
            13        29        42         57
                      16        13         15  ^
                                              acc
                                              159
                                              102

A job starts at 09:08 and finishes at 10:06. It will not be present in a 'ps' snapshot taken at 09:00 (it hasn't started yet). It will be present in the 'ps' snapshots taken at 09:15, 09:30, 09:45 and 10:00. It will not be present in the 'ps' snapshot taken at 10:15 (it finished nine minutes earlier). The 'ps' command gives a running total of the CPU seconds of a process. At 09:15 we see that it has accumulated 13 CPU seconds since it started; at 09:30 the total has risen to 29, i.e. another 16 have been added. Another 13 in the next 15 minutes and another 15 in the final quarter of an hour

What we can't tell from that is that this particular job does some very intensive processing at the end. If accounting is on, that will tell you that the job consumed 159 CPU seconds in total, and you'll be able to work out that it used 102 of them in its last six minutes.

Accounting also shows the split of CPU time between system state and user state ('ps' doesn't), the average memory occupancy of a process ('ps' doesn't) and the number of logical block I/O requests that the process made ('ps' has no I/O information of any kind).

The Athene software automatically integrates the 'ps' and UNIX accounting data into its Performance Database to provide the fullest possible picture of user/process activity. Whether using Athene or other means, this technique allows accurate analysis of machine performance, and provides the best foundation upon which to build planning models of your business.