On Sun, Mar 28, 2010 at 11:30 PM, Rainer Heilke
<rheilke at dragonhearth.com> wrote:
> Congrats on the OGB seat.

Thanks. I think...

> Good work on the tool. I haven't explored it properly yet, but it sounds
> good. I think the idea of getting everything automatically is great. We
> often find we need to examine a problem using X stats. "Oh, wait. We don't
> collect X." "Add X then, and hope the problem happens again." -- Seems a
> little backwards.

Indeed. Wouldn't it have been nice to have that data all along?
(Fishworks just accumulates everything for the same reason, I think.)

I've written a little app to summarize some of the properties of the
accumulated statistics. On my desktop development machine over
this evening:

Total kstats: 713
Total statistics: 15076
Numeric statistics: 14816
String statistics: 260
Statistics zero: 8331
Statistics Changed: 2476
Kstats changed: 262

What that tells you is that most of the data is always zero. Then about half
as much is non-zero but static. Only half of that again actually contains
interesting data. So only about 1/3 of the kstats and 1/7 of the statistics are
interesting. The others could be optimized out of the history.

(That's still several thousand potentially interesting statistics, though.)

> I think having a simple mechanism to map disks would be very good. Too many
> times, I had to do this manually to find what was really going on.

The question is how to do that. One way would be to fake up a bunch
of kstats (maybe emulate the mntinfo kstats that nfs generates) that
get saved the same way as the rest. It's definitely cheating, but it has
the advantage of being in-band, as it were.

-- 
-Peter Tribble
http://www.petertribble.co.uk/ - http://ptribble.blogspot.com/

Reply via email to