# [tex-live] strange discrepancy in running time of etex between TL2015 and TL2017

Reinhard Kotucha reinhard.kotucha at web.de
Fri Jul 28 05:36:45 CEST 2017

On 2017-07-26 at 18:45:19 -0600, Nelson H. F. Beebe wrote:

> One other point to note when comparing times of small benchmarks is
> that modern processors, especially in laptops and compute servers,
> monitor their workload, and decrease clock frequency when they
> don't have much to do.  The reason is that energy consumption is
> proportional to (clock frequency)**x, where x is usually some
> number between 2 and 3.

Dear Nelson,
the relationship between clock frequency and power consumption is
linear.  It always takes the same amount of energy to charge or
discharge a capacitor, it only matters how often you do it.

Power consumption also depends on the operating voltage and this
relationship is quadratic according to the formula

$E = \frac{1}{2} C V^2$

This is the reason why CPU core voltages are so low nowadays.

I noticed recently that modern computers reduce the clock frequency
when there is noting to do.  I assumed this already but didn't care
until I played with a small computer based on an ARM processor running
Ubuntu LTS with the real-time extensions enabled in the Linux Kernel.
The idle frequency is 200 MHz and it switches to 800 MHz within 300 µs
if loaded.  This is a feature of the Linux kernel.  Don't know whether
it's necessary to recompile the kernel in order to change the
parameters.  Maybe it's sufficient to simply poke them into the /proc
filesystem.

The machine I'm currently using runs a Gentoo system with a customized
kernel.  xosview reports a constant CPU frequency of 2493 MHz and it
seems that I wasted a lot of energy in the past.  My new machine is
running Centos 7.  The clock frequency is 800 MHz in idle state and

Hence the tests I did on my old machine are not affected by varying
CPU frequencies.  I also believe that it's acceptable for benchmarks
if it takes only 300 µs in order to switch to full speed, even on fast
machines.

Regards,
Reinhard

--
------------------------------------------------------------------
Reinhard Kotucha                            Phone: +49-511-3373112
Marschnerstr. 25
D-30167 Hannover                    mailto:reinhard.kotucha at web.de
------------------------------------------------------------------