I wanted to avoid any filesystem specifc behavior from the output.
That will be very hard. You'd need a DTrace script that breaks-out the time spent in the different layers down the IO stack, and substract the time spent in the filesystem layers from the total.
The problem remains the same: profile app run time. sys, usr, various wait's. From your reponse, it becomes clear that think I have been using the wrong term, instead of calling wait's that time is sleep. So, a breakup of sleep time among various reasons.
There's a script call "whatfor.d" in /usr/demo/dtrace that may make a starting point for you, but it does not track time, just the reason a thread was put to sleep. The problem you'll run into is most sleeps are on condition variables, so if you're thread is sleeping on IOs, be they disk IOs or network IOs, you'll see the threads sleeping on a CV. Further decomposing that into sleeps on disk IO versus network IO, and than further decomposing disk IO into time spent in the FS layers versus time spent in the driver layers, is hard. Off hand, I can't think of a good starting point to help with your first requirement. I'd like to spend some time looking at it, because it's an interesting question to answer. Unfortunately, I can not commit to spending any time on it in the next couple weeks, due to the current work queue backlog.... Thanks, /jim _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
