On Thu, May 13, 2010 at 11:52 PM, Tom Erickson <thomas.erick...@oracle.com> wrote: > Peter Tribble wrote: >> >> My plan to archive kstats using kstat -p output just got wrecked. >> The snag is that for IO statistics (and I presume timer stats, but >> they're pretty rare) the times are hrtime_t and kstat -p formats >> the output in seconds as a floating number: .. > Would it break anything if just 'kstat -p' always printed times in > nanoseconds? If so, could you add an option to kstat that forces time values > to be printed in nanoseconds?
I had a quick look at this and that doesn't look too easy. The snag is that the perl module deep in its core takes the hrtime_t, divides it by 10^9 and converts it to a float. It's not just a display tweak. So my next option is to quickly write something that largely emulates kstat -p (with the kstat type added and the IO statistic times in the underlying nanoseconds). This has the advantage that the tools I'm using to read it don't need to be modified, and doing it in C is much quicker that the perl implementation (so the measurement is less likely to perturb the system). -- -Peter Tribble http://www.petertribble.co.uk/ - http://ptribble.blogspot.com/ _______________________________________________ observability-discuss mailing list observability-discuss@opensolaris.org
