climateprediction.net home page
WU energy vs value?

WU energy vs value?

Message boards : Number crunching : WU energy vs value?
Message board moderation

To post messages, you must log in.

AuthorMessage
Venkatesh Srinivas

Send message
Joined: 7 May 17
Posts: 16
Credit: 3,480,030
RAC: 2,845
Message 56730 - Posted: 27 Aug 2017, 3:30:44 UTC
Last modified: 27 Aug 2017, 3:31:08 UTC

Anyone have any thoughts as to when (energy-wise and emissions-wise) it is worth running a CPDN workunit?

In June my computer ran a wah2_pnw25_c25q_200312_49_588 task, taking 1,077,019 CPU-seconds. Dynamic power of one core working on a CPDN WU is 30 W above idle. This is ~30,000,000 watt-seconds => ~9 kWh. Where I live average lb CO2e/mWh is 1,498, so this workunit generated ~13 lb CO2e.

At this rate, is it worth running CPDN WUs here? Or should we prefer to only run them where we are more power-efficient (FLOPS/W) or cleaner (lb CO2e/mWh)?
ID: 56730 · Report as offensive     Reply Quote
Jim1348

Send message
Joined: 15 Jan 06
Posts: 637
Credit: 26,751,529
RAC: 653
Message 56734 - Posted: 28 Aug 2017, 1:20:22 UTC - in response to Message 56730.  

In June my computer ran a wah2_pnw25_c25q_200312_49_588 task, taking 1,077,019 CPU-seconds. Dynamic power of one core working on a CPDN WU is 30 W above idle.

The longest-running pnw25 for me since the beginning of 2016 took 294,630.70 seconds on a virtual core of an i7-4770. Those cores usually take about 12 watts additional power as I recall (not counting the motherboard, etc.). I think you could use a more efficient rig.

But whether it is "worth it" is impossible to judge. How will the information be used? We don't know.
ID: 56734 · Report as offensive     Reply Quote
Jim1348

Send message
Joined: 15 Jan 06
Posts: 637
Credit: 26,751,529
RAC: 653
Message 56735 - Posted: 28 Aug 2017, 1:39:22 UTC - in response to Message 56730.  
Last modified: 28 Aug 2017, 1:40:54 UTC

And some people never return a work unit in months, or years or ever. I think we can say their contribution is not worth it. So if you are returning something useful, that is better at least.
http://climateapps2.oerc.ox.ac.uk/cpdnboinc/forum_thread.php?id=7674

Is driving your car worth it? The PC uses a lot less power, for whatever that is worth.
ID: 56735 · Report as offensive     Reply Quote
Profile Iain Inglis
Volunteer moderator

Send message
Joined: 16 Jan 10
Posts: 1079
Credit: 6,904,878
RAC: 6,593
Message 56741 - Posted: 28 Aug 2017, 19:58:55 UTC

Those are interesting numbers, Venkatesh, and concentrate the mind.

I would invite you to consider the following counter-arguments:

1. The energy cost of running a CPDN model perhaps supposes that the purpose of the project is to encourage mitigation of climate change. It would be a fascinating if difficult exercise for someone to attempt to link the papers published as a result of research here to improved mitigation (i.e. x kg CO2 emitted for y kg saved). In general this will not be known until significant mitigation takes place, even if attributing effects to causes were possible - what about Al Gore and a million other people and initiatives?

2. Some of the recent work here concerns regional-scale climate impacts, which would more naturally be characterised by avoided financial costs. Such work may also contribute indirectly to mitigation, since it brings home to people how they will be affected by heat waves, droughts, floods, sea level rise etc. That sense of personal involvement may therefore contribute not only to adaptations (such as flood defences) but to a desire to avoid the impacts by reducing greenhouse gas emissions (i.e. mitigation). Would the generation of CO2 now be justified by the saving of money later?

3. Climate science is an interesting subject in its own right and would be investigated whether or not it was of topical interest because of climate change. Nowadays that investigation would involve significant computing resources, probably supercomputers - and the energy cost of that is in part replaced by the use of personal computers here.

4. The energy argument you have raised appears from time to time on this message board and I always wonder why only efforts to understand the climate should be subject to an energy test. Why not computer games? or TV? or sports? or packaging? etc? etc?
ID: 56741 · Report as offensive     Reply Quote
Profile tullio

Send message
Joined: 6 Aug 04
Posts: 264
Credit: 965,476
RAC: 0
Message 56742 - Posted: 29 Aug 2017, 10:07:28 UTC - in response to Message 56741.  

I am running also SETI@home, Einstein@home, LHC@home. Nobody doubts about the validity of those. I could be watching "Game of thrones" on my TV set, which works as a monitor.
Tullio
ID: 56742 · Report as offensive     Reply Quote
WB8ILI

Send message
Joined: 1 Sep 04
Posts: 161
Credit: 81,421,805
RAC: 1,225
Message 56743 - Posted: 29 Aug 2017, 12:23:05 UTC

Assuming the pool of human knowledge is being increased by running these models, I would suggest that the energy cost is worth it.

It doesn't matter if one is studying climate change or the sex habits of a mosquito, the value of increased knowledge cannot be calculated.
ID: 56743 · Report as offensive     Reply Quote
Venkatesh Srinivas

Send message
Joined: 7 May 17
Posts: 16
Credit: 3,480,030
RAC: 2,845
Message 56744 - Posted: 29 Aug 2017, 12:27:36 UTC

Sorry, I asked my question poorly earlier - I wasn't asking if CPDN is worth running overall.

I was trying to ask if CPDN is worth running at a particular FLOPS/W (when a particular WU runs for a 1 million seconds at +30 W over baseline). A poster above noted that his workstation was able to complete a similar WU in 300,000 seconds at +12 W, nearly an order of magnitude fewer watt-seconds. (Do all pnw25 WUs have comparable amounts of work?)

If we have some kind of idea of the value of one WU, we can decide between running CPDN locally or (say) renting time on more power efficient computers. Even if we don't have a dollar-based value function, an idea of how many trajectories go into a given project and what avg performance of computers delivering results would help with those kinds of decisions.
ID: 56744 · Report as offensive     Reply Quote
Jim1348

Send message
Joined: 15 Jan 06
Posts: 637
Credit: 26,751,529
RAC: 653
Message 56746 - Posted: 29 Aug 2017, 13:48:16 UTC - in response to Message 56744.  

I was trying to ask if CPDN is worth running at a particular FLOPS/W (when a particular WU runs for a 1 million seconds at +30 W over baseline). A poster above noted that his workstation was able to complete a similar WU in 300,000 seconds at +12 W, nearly an order of magnitude fewer watt-seconds. (Do all pnw25 WUs have comparable amounts of work?)

That was the worst-case (longest-running) scenario for me. The average PNW has been about 250,000 seconds since the beginning of 2016. But I was wrong on the CPU. It was not run on my i7-4770 (Haswell) but rather an i7-3770 (Ivy Bridge), which uses a little less power; certainly less than 10 watts per virtual core. I think your order-of-magnitude estimate is correct.

That is not to say your present setup is not useful, but efficiency could be improved if you want to. I will be building a Ryzen 1700 machine shortly, which should improve it more.
ID: 56746 · Report as offensive     Reply Quote
Profile geophi
Volunteer moderator

Send message
Joined: 7 Aug 04
Posts: 2167
Credit: 64,403,322
RAC: 5,085
Message 56749 - Posted: 30 Aug 2017, 2:15:18 UTC - in response to Message 56746.  

I will be building a Ryzen 1700 machine shortly, which should improve it more.

Not wanting to discourage you, but if you are going to install Linux on it, you may want to wait. Ryzen problems with some heavy workloads in Linux have been documented. There are several news stories at Phoronix on this. I have a Ryzen 1500X and it is having troubles running more than a couple tasks at a time without segfaults (this is on known good tasks that should complete on a stable machine). This Ryzen can run Prime95 for 48 hours on 8 cores without any errors, but load up 4 cpdn tasks and within 18 to 24 hours, one or two will segfault. I am going to try some WCG tasks to see if the problems are cpdn specific.

No problems like this have been reported on Windows at this time.

Threadripper and new steppings of the Ryzens have fixed this, but I'm not sure about whether some fix to the Linux kernel can workaround the problem on the initial stepping Ryzens that are still being sold. How will one know if they have a new, fixed stepping when they buy it?

Article on Threadripper and new stepping Ryzens being fixed, but has links to original articles on the problems previously encountered in Linux under some heavy loads.
ID: 56749 · Report as offensive     Reply Quote
Jim1348

Send message
Joined: 15 Jan 06
Posts: 637
Credit: 26,751,529
RAC: 653
Message 56750 - Posted: 30 Aug 2017, 4:56:09 UTC - in response to Message 56749.  
Last modified: 30 Aug 2017, 5:32:57 UTC

Thanks. I am aware of the errors, but that machine won't be so heavily loaded. It is primarily for Rosetta and WCG. I have not heard of problems there, though there is no way of knowing how a particular chip and motherboard will behave until you try. I was going to go with Intel Coffee Lake, but they have delayed it yet again. Who knows when they will get the bugs out of that one also.

Also, the motherboard plays some part. As you know, memory compatibility has been a big problem, so I chose the best I could find at a reasonable price, an ASRock Fatal1ty X370 GAMING X. I am not a gamer, but I wanted the 12-phase (8+4) power, with good VRM heatsinks, for stability. And I will use only DDR4 on the qualified vendors list. I never overclock, so something will work on it. It is just a question of what. Maybe a few CPDN will run also. Actually, they would be a good test. I will start out with 16 of them and see what happens (Ubuntu 16.04, now that you can install the 32-bit libraries).
ID: 56750 · Report as offensive     Reply Quote
Jim1348

Send message
Joined: 15 Jan 06
Posts: 637
Credit: 26,751,529
RAC: 653
Message 56757 - Posted: 1 Sep 2017, 1:44:57 UTC - in response to Message 56750.  
Last modified: 1 Sep 2017, 1:55:35 UTC

This is a bit OT, and perhaps more for a hardware forum, but since the issue has been raised, it appears that AMD has fixed the segfault problem on the newer Ryzens.
https://www.reddit.com/r/Amd/comments/6ubmd1/ryzen_compilation_segfaults_positive_rma/

I always wait six months after a CPU is actually released before buying one. It was released on March 2, so Labor Day sales may work for me in the U.S. But I could wait until Ubuntu 17.10 is released in October. I am in no great hurry.
ID: 56757 · Report as offensive     Reply Quote
Profile geophi
Volunteer moderator

Send message
Joined: 7 Aug 04
Posts: 2167
Credit: 64,403,322
RAC: 5,085
Message 56758 - Posted: 1 Sep 2017, 2:04:44 UTC - in response to Message 56757.  

To continue OT, the article I linked to tells how to tell whether it is a new stepping or not, contained in the date of manufacture on the heatspreader. Unfortunately, unless one is looking at the heatspreader itself, it will be hard to know whether you are getting a fixed one or not. BTW, I've been running WCG for 24 hours on all 8 virtual processors with no errors. Perhaps a cpdn thing.
ID: 56758 · Report as offensive     Reply Quote

Message boards : Number crunching : WU energy vs value?

©2024 climateprediction.net