LOAD_FREQ (4*HZ+61) avoids loadavg Moire

LOAD_FREQ (4*HZ+61) avoids loadavg Moire

[Klaus Ripke]

LOAD_FREQ is (5*HZ+1) to avoid high load average when idle:
http://kerneltrap.org/mailarchive/linux-kernel/2007/10/3/328568

I suggest (4*HZ+61) for a better distribution.

With some seconds based load (like SSL heartbeats)
and LOAD_FREQ at (5*HZ+1) I see Moire patterns like inverse sawtooth,
since 2 or 3 probes hit the jobs (load increases quickly),
followed by several probes missing it.


A 4.61 sec interval gives optimal distribution over when within a
second a probe is taken, as .61 is close to golden ratio phi 1.618...
(test in http://ripke.com/goldenratio.c).

12*4.61 = 55.32 secs is still close to a minute,
and 13*4.61=59.93 is even closer than the current 12*5.01=60.12
(with exponents EXP_x adjusted to a ratio of 13 instead of 12).


best
Klaus

Re: LOAD_FREQ (4*HZ+61) avoids loadavg Moire

[Clemens Ladisch]

Klaus Ripke wrote:
> With some seconds based load (like SSL heartbeats)
> and LOAD_FREQ at (5*HZ+1) I see Moire patterns like inverse sawtooth,

Thanks for researching this, but I don't see them because I'm too lazy
to reproduce this.  Could you please publish a colorful screenshot
somewhere?

> I suggest (4*HZ+61) for a better distribution.

It appears your attached patch got lost somehow.  ;-)


Regards,
Clemens

Re: LOAD_FREQ (4*HZ+61) avoids loadavg Moire

[Klaus Ripke]

On Mon, Nov 14, 2011 at 09:10:44AM +0100, Clemens Ladisch wrote:
> Klaus Ripke wrote:
> > With some seconds based load (like SSL heartbeats)
> > and LOAD_FREQ at (5*HZ+1) I see Moire patterns like inverse sawtooth,
> 
> Thanks for researching this, but I don't see them because I'm too lazy
> to reproduce this.  Could you please publish a colorful screenshot
> somewhere?

there you go http://ripke.com/loadavg/moire


best

Re: LOAD_FREQ (4*HZ+61) avoids loadavg Moire

[Alan Cox]

On Wed, 16 Nov 2011 12:03:30 +0100
Klaus Ripke <klaus@ripke.com> wrote:

> On Mon, Nov 14, 2011 at 09:10:44AM +0100, Clemens Ladisch wrote:
> > Klaus Ripke wrote:
> > > With some seconds based load (like SSL heartbeats)
> > > and LOAD_FREQ at (5*HZ+1) I see Moire patterns like inverse sawtooth,
> > 
> > Thanks for researching this, but I don't see them because I'm too lazy
> > to reproduce this.  Could you please publish a colorful screenshot
> > somewhere?
> 
> there you go http://ripke.com/loadavg/moire

And for any load you pick you'll be able to generate patterns and
inaccurate load average data unless you actually count used cycles via
the perf counters and such on processors.

Shannon says we can't win. Changing the computation may well change load
values other users use in order to predict and balance workloads so while
it has clear limits futzing with it to make one workload look better is
simply going to break stuff for other people.

Alan