On Wed, 23 Nov 2005 13:39:14 -0500, David Andrews <[EMAIL PROTECTED]> wrote:
>On Wed, 2005-11-23 at 11:25 -0600, Tom Schmidt wrote: >> Count the WTOs in a syslog for an estimate of the WTO >> cost per subsystem. > >Taking Ed's estimate of three instructions per inactive SSCVT, and >assuming 10M WTOs per day, and ten inactive subsystems... that's roughly >a CPU second per day spent cycling down the chain. > >There must be bigger fish to fry. Perhaps true, but the issue isn't entirely one of the amount of CPU seconds per DAY. A large part of performance has to do with the nonuniform arrival rate of the instructions in question. Specifically, the activity doesn't spread itself evenly thoughout the day -- it usually happens in clumps or clusters. Usually when you are already busy (and in these cases generally when lots of tasks are terminating with lots of messages being issued). So having the extra "few" instructions serves to make a bad situation worse. I'm generally opposed to making bad things worse - I believe I am paid to make bad things better. When you reduce the size of the clump (i.e., the height on the graph) then you give yourself more headroom. And things get better (even if it isn't obvious when things aren't bad). -- Tom Schmidt Madison, WI ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html
