Sure, I pretty much agree with your analysis. The old code calculated an average lifespan for subfacets over the entire life time of the switch. I suppose I would argue that the old metric isn't particularly useful. Who cares if 2 days ago most of the subfacets lasted 300 milliseconds? When I'm looking at the switch, I want to know at this moment (give or take a minute), are the flows we're looking at short-lived or long lived.
At any rate, the commit message should have done a better job of pointing the change out. I didn't really design the feature in the first place, so perhaps the original behavior is what was intended. I dont' feel all that strongly about it to be honest. Just seemed a bit weird to me. Perhaps Justin will have an opinion since he put in the original feature request? Ethan On Thu, May 30, 2013 at 7:38 PM, Ben Pfaff <b...@nicira.com> wrote: > On Thu, May 30, 2013 at 02:45:12PM -0700, Ethan Jackson wrote: >> Along with some minor minor cleanups, this patch simplifies the >> calculation of the average subfacet count, and average subfacet >> lifespan. They are both calculated as exponentially weighted >> moving averages updated immediately after statistics are pulled >> from the datapath. This both simplifies the code, and makes the >> statistics produced more useful. Specifically, with the average >> total subfacet lifespan, subfacets which are still in the datapath >> now count towards that average. > > I think (without really looking at it) that the previous code for > average subfacet lifespan calculated a genuine average over every > subfacet that ever existed. The new version weights currently > existing subfacets the highest (but also gives them some additional > credit for all the time they existed previously--is that kosher?). > Subfacets that have expired receive exponentially decaying weight, so > that a subfacet that disappeared 10 seconds ago essentially disappears > from the average too. That's a pretty different (and much > shorter-term) statistic. > > If you want to include subfacets currently in the datapath, then you > could fold those in with a loop similar to the one that you added in > this patch. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
