I expect a difference between the CPU usage recorded by SAR and that recorded by UNIX process accounting, but how big should it reasonably be?

If you have been measuring the system for a reasonably long period (say 1 hour) then the system v accounting discrepancy should be no more than 5%. This is on the assumption that short-duration jobs have been recorded by the 'accton' command, and long-running jobs have been measured using 'ps' snapshots. If the difference is larger than that, there are one or two possibilities. There could be a large kernel time element, involving kernel processes not directly related to visible UNIX or user loggable processes. Processor management in a large multi-processor machine is one example. On a very fast machine, the usual culprits are very short-lived processes. UNIX accounting only reports the CPU usage of a process to two decimal places, i.e. hundredths of a second. On today's high speed CPU's, processes can run and terminate in milliseconds, and these will show up as 0.00 seconds in the accounting. We saw one site where in one day, the same process was executed 930,000 times, but each execution was averaging 5 milliseconds. This generated a loss of recorded CPU usage of 4,650 seconds over the day.