10-29-2009 04:23 PM
I have a sampling activity for which I'm trying to estimate several ratios. When I ran the activity, vtune split it into multiple runs (7 in all). It appears that CPU_CLK_UNHALTED.CORE is only sampled for the first run even though ratios computed from other runs require CPU_CLK_UNHALTED.CORE in the divsor. Does that make sense? What happens if two runs take a different number of clockticks, wouldn't that skew the ratio?
10-29-2009 05:15 PM
The assumption is that the multiple runs are directly comparable, which certainly breaks down when there are competing processes running on the system. When there aren't many samples in a region of interest (most samples occurring elsewhere), there can be quite a bit of non-repeatibility entering into the comparison of the multiple runs. So, it may be helpful to set up runs to increase the time spent in a region of interest and reduce the magnitude of the problem you appear to be pointing out. You may get more insight, as you imply, by setting up individual runs which count just clock ticks and few enough other events to fit in a single run.