Arthur, In my experience the TIMEUSED macro is useful chiefly for timing the CP usage of entire routines, which may of course be short ones, not individual instructions.
Care is required in using it. The two successive TIMEUSED macros employed to obtain a difference/measurement should be so placed that I/O activity and, more generally, SRB-mode activity of any kind does not occur between them (because SRB-mode CP time is not included in TIMEUSED values). My answer to the question o How precise are the measures provided by TIMEUSED? needs to be qualified. Statisticians distinguish precision and bias. Precision has to do with reproducibility, with the extent to which the variance of a set of measurements is small. In this sense TIMEUSED values are very precise, have low variance. Accuracy/bias issues, i.e., those of systematic under- or over-estimation, are more difficult to evaluate. My guess is that these differences are highly accurate, but this guess is based upon the fact that they behave as I should expect them to behave when I change a routine the CP time of which is being measured The answer to the question o Do these measures take into account the delays caused by cache misses? is simple. Yes. The answer to your last question o . . . does one have to expect effects of some kind of Uncertainty principle (i.e. the measurement algorithm, could it influence the result)? is only a little more complicated. Residence/elapsed time is of course affected by the use of TIMEUSED, but as many as 15 successive cheek-by-jowl (conceptiually vacuous) TIMEUSED macro instructions affect the results they yield negligibly. (It may indeed be that TIMEUSED itself uses SRB-mode processing. I should have done so if I had implemented it.) I hope these 'answers', which are not definitive, are of some help. I have many unanswered questions of my own about TIMEUSED; but it is, I think, a good piece of work, eminently usable and free of nasty surprises. If you have further, more detailed questions that are not of general interest send them to me directly. John Gilmore, Ashland, MA 01721 - USA ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO IBM-MAIN
