A farewell to set_fs()?

Posted Jun 9, 2017 14:43 UTC (Fri) by mkbosmans (subscriber, #65556)
In reply to: A farewell to set_fs()? by fratti
Parent article: A farewell to set_fs()?

Let's try to run the numbers.

Assuming:
- 2e9 computers, phones, etc. on the world run Linux
- They are active 10% of the time
- On average they run 3e9 instructions / second
- When active, there are 1e4 syscalls / second
- When active, 30 W of power is used

Then:
- 5 instructions of overhead for each syscall results in a 5 * 1e4 / 3e9 * 30 = 0.5 mW extra power when a computer is active.
- Globally, this means an increased power usage of 0.5mW * 10% * 2e9 = 100 kW
- On a yearly basis, this amounts to 876000 kWh, which equals to about 700 tons of carbon emissions.

According to [1], the smallest countries, like Tuvalu, emit 3000 tons per year. So, although not quite there, it is reasonably close.

[1] http://cdiac.ornl.gov/trends/emis/top2014.tot

A farewell to set_fs()?

Posted Jun 9, 2017 23:12 UTC (Fri) by cjr (guest, #88606) [Link] (1 responses)

That's pretty interesting. The only figure I would take issue with is the 30W power consumption. Assuming most of those 2e9 devices are phones, the power consumption is probably much, much less than 30W. It's probably on the order of 1W or less.

All the information I could find in a quick search is a bit outdated, but here it is anyways:

This paper has some analysis of mobile phone power consumption:
https://www.usenix.org/legacy/event/atc10/tech/full_paper...

This article from Qualcomm (2013) has some stats based on the above article:
https://developer.qualcomm.com/blog/mobile-apps-and-power...

This article (behind an obtrusive advertisement FYI) says it takes ~1kWh per year to run a mobile phone:
https://www.forbes.com/sites/christopherhelman/2013/09/07...

Chris

A farewell to set_fs()?

Posted Jun 10, 2017 16:09 UTC (Sat) by mkbosmans (subscriber, #65556) [Link]

I based the number of 2 billion Linux installations on an article [1] dated 2008 reporting that there the number of personal computers passed the 1 billion mark.

So the number of desktops, laptops and servers (i.e. computers with a 30W or greater power usage) combined will probably be more than 2 billion right now. Of course not all of them run Linux.

There have been sold 1.6 billion android phones to date [2]. Excluding discarded phones and including routers and other embedded devices brings the total of active low-power Linux devices probably also in the range of 1-2 billion.

Anyway, this whole exercise was more to see whether the power savings could even close to the power consumption of a small country. Even if a lot of guesses are wrong, I don't think I'm more than a factor 1000x off.

[1] http://www.reuters.com/article/us-computers-statistics-id...
[2] http://www.statisticbrain.com/android-phone-statistics/

A farewell to set_fs()?

Posted Jun 9, 2017 23:37 UTC (Fri) by cjr (guest, #88606) [Link] (1 responses)

As an addition to my previous comment:

Assuming the 30W power consumption was correct, are you sure you have the math right? I get the following:

Note .05mW rather than .5mW:
5 * 1e4 / 3e9 * 30 = 0.00005 = 0.05 mW
0.05mW * 10% * 2e9 = 10 kW
10kW * 24 * 365 = 87600 kWh

After going through this exercise, I also found that I was surprised if even an average desktop CPU would consume .05mW to execute 5000 instructions. So I looked on Wikipedia [1] to find a Core-i7 940XM, which uses 55W at 2.13GHz. 5000 instructions in that case would take ~.129mW, which means indeed it can take more power than I thought to execute those 5000 instructions. But even if all those 2e9 devices were running that hungry Core-i7, they would still be consuming roughly 4x less power than the original figures.

Thanks for provoking an interesting discussion anyways. Also please correct me if I have gotten anything wrong above.

[1] https://en.wikipedia.org/wiki/List_of_CPU_power_dissipati...

Chris

A farewell to set_fs()?

Posted Jun 10, 2017 16:13 UTC (Sat) by mkbosmans (subscriber, #65556) [Link]

Please note that I estimated 10.000 syscalls per second.
I did this by firing up a browser under strace and looking up a website. That was 45.000 syscalls in 5.5 seconds.

Of course this is a really high estimate, because even when a user is actively using his computer, the average number of syscalls/second will be much lower.

Another big simplification is 1 instruction == 1 cycle. That can certainly be more or less, depending on the specific instructions and other context. But again, this whole exercise was meant as a https://en.wikipedia.org/wiki/Fermi_problem.