[ruby-core:33197] Re: Power consumption of timer thread

Excerpts from SASADA Koichi's message of 2010-11-14 16:34:20 -0500:
> Hi,
> 
> * Abstract
> 
> Some people pointed out that 1.9 timer thread introduce power
> consumption issue.  However, nobody show the quantitative evaluation
> (*1).  We measured the power consumption using watt checker on the Intel
> Core i5 processor.  Our measurements  show that 1 idle ruby process only
> consumes few power (0.2-0.3W), not a 0 power (idle 1.8 process consumes
> zero power).  However over 100 idle process consume some watts.
> 
> *1: powertop only shows the interrupt number.
>     Not a real power consumption.

Thank you for spending time on this issue.

If you're running an unplugged laptop, powertop will show actual power
consumption from the battery.  But, your power meter results are also
going to be accurate.

> 
> * Background
> 
> Omitted.  Read the related articles :)
> 
> Ah, I want to add information that I need to talk about this scheduler
> modification at a conference next year :)  This is because I make a such
> measurement and write a such article :)
> 
> 
> * Evaluation
> 
> ** Evaluation environment
> 
> - H/W
> -- Intel Core i5
> --- 4 cores
> -- 4GB memory
> 
> - S/W
> -- Ubuntu 10.10 Server 64bit (no configuration changed)
> -- ruby 1.9.2p0 (2010-08-18 revision 29036) [x86_64-linux]
> 
> - measurement software: fork.rb (attached)
> -- phase 1. Run empty loop 3 seconds.
> -- phase 2. Fork n ruby idle processes and run 60 seconds.
>              n is 0 to 256 (0 and 2 ** i, i is from 0 to 8).
> -- phase 3. Kill n forked process and back to phase 2.
> -- Of course, fork.rb is run by ruby 1.8 :)
> 
> - Power consumption measured by "Watts up? PRO"
> -- https://www.wattsupmeters.com/secure/products.php?pn=0
> 
> ** Evaluation result
> 
> See attached graph fig01.png and raw data power.xlsx.
> X-axis shows time.  X-axis shows power consumption (Watt).
> Red line is power consumption.
> Blue line is 52.1 line (0 process line).
> 
> Peak power consumption over about 80 watts means 3 seconds empty loop
> (fork.rb phase 1).  We can recognize the n with this peak.
> 
> Running only 1 idle process, we can't recognize increasing power
> consumption on the fig01.  Results show that 0.2-0.3W power consumption.
>  Over 4 idle process, we can see the increasing of power consumption on
> the fig01.
> 
> 
> * Conclusion
> 
> By our measurements, we can recognize the increasing power consumption
> on the Intel core i5 processor.  1 process consumes 0.2-0.3W and 256
> processes consume about 3 watts.
> 
> 
> BTW, I want to know how many idle process available on pragmatic usage.
> 
> If you have other measurement methods, please teach me.  And any other
> comments are welcome.

The power consumption is only really increased if the machine is
otherwise idle.  If the machine is able to power down a CPU, waking that
CPU up for the timer thread is fairly expensive.  If the CPU is already
awake doing other tasks, the timer thread has a fairly small cost.

-chris