On Tue, 12 Sep 2006 10:51:02 -0400, Thor Lancelot Simon <tls@rek.tjls.com>
wrote:

> On Tue, Sep 12, 2006 at 09:46:24AM -0400, Steven M. Bellovin wrote:
> > 
> > It's not clear to me that taking the average is worthwhile -- I suspect
> > that at times that it matters, the statistical variant I suggested will
> > work.  But basically, we all agree -- don't worry about the underlying
> > device or the file system layout properties; just look for writes that are
> > starting to take longer than they "should" when things are busy.
> 
> Yes, but this can tend to penalize those process whose writes take longer
> than they "should" because of some _other_ process' pathological behavior.
> 
> This is why I suggested that what you want to do is penalize those processes
> for whom writes that "take longer than they should" are the preponderance of
> writes -- or, considered differently, to penalize processes according to the
> probability that their writes occur when the average latency across all
> writers (which is an easier statistic to measure anyway) happens to be
> increasing.  This nails the pathological "scatter write" patterns with a
> penalty while minimizing the penalty to the writers who they disrupt, I
> think.
> 
I think we're in violent agreement.  As you say, we should "penalize
processes according to the probability that their writes occur when the
average latency across all writers (which is an easier statistic to
measure anyway) happens to be increasing."  I agree -- but I'm asserting
that the ancient Unix time accounting scheme -- bill whatever process is
active when the timer ticks -- will work here.  A process that issues few
writes during that period will be penalized, but statistically very
slightly; ones that issue a lot of writes will statistically incur more of
a penalty.


		--Steven M. Bellovin, http://www.cs.columbia.edu/~smb