I do know they're writing a 128-bit kernel, which you could call forward-thinking but considering we're still in the transition from 32 to 64 bits, I think it's a bit silly (hell, modern processors can't even address all 64 bits of memory yet).
If it's not an outright red herring that linked in profile (job ad?) has got to be messed up/taken badly out of context. Quite simply there's no such thing as a 128bit integer x86 CPU on the market or being talked up in anyones future architecture slides; nor is there a need for one in anything other than a few specialized tasks. The only explanation that makes any sense is that they'll be making use of 128 bit SSEn instructions; probably in the video pipeline somewhere.
If anyone wants any further embellishment about my budget crunching and upgrading experiences, just ask away.
You've included lots of detail, and are getting admirably high RAC per capital dollar. But I don't spot that you've mentioned power consumption considerations (save only the implications of the power supplies you use).
My impression is that even for people pushing higher toward the bleeding edge in price than you that lifetime power consumption costs are appreciable, and properly considered would cause many to back down to lower overclocks and overvolts than the ones they choose (myself, I choose stock clock and undervolt for this reason).
All other things being equal, your work more toward the value end of the spectrum would be expected to make the relative cost contribution of lifetime amortized power even higher, and thus the optimum overclock overvolt even less.
Is there some reason power matters less to you than it might to some others (someone else paying?...) Even if you don't care for your own circumstance, the (considerable) value of your examples to others would be enhanced with even a little power information. I'd be happy to purchase a power meter and ship it to you if you think you might occasionally use it and report, for example, RAC/watt at the system level. What is local voltage/frequency, and would mechanical socket adapters be acceptable to you?
Great report, Gary. But it really would be interesting to know about power cosumption of your farm. I would be glad to build something like this - I have enough electricity for only $10 per month :) But all I have now is only a bunch of Celeron's P-III and a lot of pentium mobos. Most of others are not mine - they are in testing after repairments and soon their owners will get them back.
If anyone wants any further embellishment about my budget crunching and upgrading experiences, just ask away.
You've included lots of detail, and are getting admirably high RAC per capital dollar.
Hi Peter, very nice to hear from you.
Yes, you are quite correct to highlight capital dollar. I, most assuredly, should have said, "budget hardware crunching", which is all I was actually trying to talk about.
Quote:
But I don't spot that you've mentioned power consumption considerations (save only the implications of the power supplies you use).
OK, I did mention the particular Seasonic model and those interested could easily google it. However, having satisfied myself that I would be using a PSU with good efficiency capable of meeting the expected load, I didn't choose to go to the trouble of purchasing a power measuring device simply because I reasoned that there really wouldn't be much I could do to lower power consumption for the conditions I knew I would be using. Let me explain a bit further.
Budget motherboards don't usually come with full voltage control options. If they have those options at all, it's usually for a modest voltage increase. I'm quite aware of the potential power savings from undervolting but the boards I looked at didn't seem to allow it.
What I was trying to do is a bit like planning to drive from LA to New York. First of all I need to choose a vehicle. I could go for a limo or I could choose the cheapest budget Asian vehicle du jour. I reasoned that both would get me there with the level of comfort and efficiency that I desired. So the choice was a no brainer. Next I had to figure out whether I wanted to travel at 70MPH or perhaps more sedately at 45MPH. I could save a lot on fuel (with both vehicles) if I could stick rigidly to 45MPH max. Like most other drivers on the roads these days, I knew that I just wouldn't be able to do that, no matter how much fuel costs I could save.
So I made the decision very early that I was in a hurry to help the project reach its goals and that I wanted to contribute more rather than less even if it meant a higher fuel bill. I regard the level of voltage and frequency increase that I use as being 'relatively modest' compared with what might be the case from using a more expensive board with full voltage options and something better than stock air cooling.
So, in short, electricity costs (though painful) would be something I was prepared to endure.
Quote:
My impression is that even for people pushing higher toward the bleeding edge in price than you that lifetime power consumption costs are appreciable, and properly considered would cause many to back down to lower overclocks and overvolts than the ones they choose (myself, I choose stock clock and undervolt for this reason).
I'm not sure I understand you here. I assume that the more expensive, more fully optioned boards will consume more power at idle than a minimum featured budget board. I also assume that I'll use less power on the budget board for any given frequency/voltage combo. The more expensive board may be able to win if I was prepared to run at stock frequency with the lowest possible volts. Since I don't wish to run at stock, it's not an option for me.
Quote:
All other things being equal, your work more toward the value end of the spectrum would be expected to make the relative cost contribution of lifetime amortized power even higher, and thus the optimum overclock overvolt even less.
Are you saying that if I spend $1000 on capital costs and $5000 on lifetime running costs, my relative power costs are higher than if I had spent $2000 on capital costs and $4500 on lifetime running costs? Also, I could do even better by lowering overclock/overvolt so that I got a bit more reduction in running costs? Maybe I could even get the lifetime power costs below $4000 so that I actually saved a bit, overall. If this is what you are saying then I fully agree with you. The only problem is that I'm not prepared to accept the slower pace. I quite like the idea of leaving the voltage at stock and just increasing frequency as much as possible, without compromising stability. I pay the frequency penalty but not a voltage one.
Quote:
Is there some reason power matters less to you than it might to some others (someone else paying?...)
I (or a close family member) pay for every watt used. I'm quite prepared to do that. Let me suggest another analogy. If you increase the price of a pack of cigarettes by an order of magnitude, would everybody stop smoking? Some people are prepared to pay (whatever, almost) for their addiction. I'm prepared to pay for mine :-). I don't smoke, drink or gamble. I'm happy to put those (and other) savings towards electricity. My next project will be to install a mini power station on my roof. We get lots of sunlight in sub-tropical Brisbane :-). If anyone has a viable design, I'm all ears! :-).
Quote:
Even if you don't care for your own circumstance, the (considerable) value of your examples to others would be enhanced with even a little power information.
OK, thanks for the suggestion. I don't currently own a power meter but I'll look into it. I believe they are readily available locally.
Quote:
I'd be happy to purchase a power meter and ship it to you if you think you might occasionally use it and report, for example, RAC/watt at the system level. What is local voltage/frequency, and would mechanical socket adapters be acceptable to you?
Thank you, most sincerely, for your kind offer but I couldn't possibly accept it. Purchasing one hasn't been high on my list of priorities but it just got bumped (a bit) :-).
My impression is that even for people pushing higher toward the bleeding edge in price than you that lifetime power consumption costs are appreciable, and properly considered would cause many to back down to lower overclocks and overvolts than the ones they choose (myself, I choose stock clock and undervolt for this reason).
I'm not sure I understand you here. I assume that the more expensive, more fully optioned boards will consume more power at idle than a minimum featured budget board. I also assume that I'll use less power on the budget board for any given frequency/voltage combo. The more expensive board may be able to win if I was prepared to run at stock frequency with the lowest possible volts. Since I don't wish to run at stock, it's not an option for me.
Quote:
All other things being equal, your work more toward the value end of the spectrum would be expected to make the relative cost contribution of lifetime amortized power even higher, and thus the optimum overclock overvolt even less.
Are you saying that if I spend $1000 on capital costs and $5000 on lifetime running costs, my relative power costs are higher than if I had spent $2000 on capital costs and $4500 on lifetime running costs?
While businesses often separate capital and operating budgets so thoroughly that they inhibit employees from being sensible, from a pure economics view one would discount the expected future cost stream associated with operating costs over lifetime and add it to the capital cost in comparing alternatives.
In other words, I think of considering the power tradeoffs in cruncher system decisions as just another aspect of the overall quest for best bang for the buck. From a pure cost point of view, for example, this is the sole reason that continuing to run a long-since purchased (and thus free in terms of capital today) antique cruncher is a bad idea, solely because the RAC/watt is so dramatically below modern systems, and that at most jurisdiction's power pricing, that cost is material in crunching.
I'm really not much questioning the wisdom of your choices--most of them look very good to my casual eye. On the other hand as a guide to others, I suspect a stronger inclusion of power cost would probably have made clear to readers that continued operation of your Tualatin fleet was actually well on the way to being big drag on your output per unit cost. Your post to my eye instead emphasizes the opportunity that upgrading them posed to provide much faster harder at a moderate capital expenditure. A balanced view would consider both aspects at once.
Of course, for many of us, it really is not the case of deploying a fixed amount of cash for maximum results--we have floorspace, spousal approval, emotional attachment to machines, and a host of other things pushing us of that optimum. Nothing wrong with any of that--I'm just a fan of rational well-informed choices.
To get back to the single decision of yours that proper consideration of power costs might most likely put in doubt--the overclocking, the possibility would be that consideration of the power cost might lead one to prefer to deploy some of the present discounted value of the future extra cost for power in the form of more machines, or possibly increased capability in some form on the same machines. This might actually be a nonsense in your case, for reasons ranging from rack space to real box availability to a difference in your practical view of present vs. future expenditures.
While businesses often separate capital and operating budgets so thoroughly that they inhibit employees from being sensible, from a pure economics view one would discount the expected future cost stream associated with operating costs over lifetime and add it to the capital cost in comparing alternatives.
In other words, I think of considering the power tradeoffs in cruncher system decisions as just another aspect of the overall quest for best bang for the buck. From a pure cost point of view, for example, this is the sole reason that continuing to run a long-since purchased (and thus free in terms of capital today) antique cruncher is a bad idea, solely because the RAC/watt is so dramatically below modern systems, and that at most jurisdiction's power pricing, that cost is material in crunching.
I'm really not much questioning the wisdom of your choices--most of them look very good to my casual eye. On the other hand as a guide to others, I suspect a stronger inclusion of power cost would probably have made clear to readers that continued operation of your Tualatin fleet was actually well on the way to being big drag on your output per unit cost. Your post to my eye instead emphasizes the opportunity that upgrading them posed to provide much faster harder at a moderate capital expenditure. A balanced view would consider both aspects at once.
Of course, for many of us, it really is not the case of deploying a fixed amount of cash for maximum results--we have floorspace, spousal approval, emotional attachment to machines, and a host of other things pushing us of that optimum. Nothing wrong with any of that--I'm just a fan of rational well-informed choices.
To get back to the single decision of yours that proper consideration of power costs might most likely put in doubt--the overclocking, the possibility would be that consideration of the power cost might lead one to prefer to deploy some of the present discounted value of the future extra cost for power in the form of more machines, or possibly increased capability in some form on the same machines. This might actually be a nonsense in your case, for reasons ranging from rack space to real box availability to a difference in your practical view of present vs. future expenditures.
G'day Pete. :-)
First class analysis! Lifetime analysis is often more accurate, at least more revealing. In the computer hardware world capital depreciates so rapidly ( effectively Moore's Law in accounting clothes ), whereas in the power supply world ( energised electrons ) the cost trend is nearly always upwards. So rated against a fixed output ( assuming RAC is some sufficiently and relatively slowly varying 'standard loaf of bread' ) the running costs will always eventually dominate. And probably sooner than you think ... :-)
[ This is a 'what's it worth on a given day' type approach - rather than historical accumulation of how much have has already been committed. This avoids the 'Concorde Fallacy' or as the Irish say "don't throw good money after bad". ]
Cheers, Mike.
( edit ) And Gary has scored two ABP's to date! :-)
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
To get back to the single decision of yours that proper consideration of power costs might most likely put in doubt--the overclocking, the possibility would be that consideration of the power cost might lead one to prefer to deploy some of the present discounted value of the future extra cost for power in the form of more machines, or possibly increased capability in some form on the same machines. This might actually be a nonsense in your case, for reasons ranging from rack space to real box availability to a difference in your practical view of present vs. future expenditures.
Well, you'll be happy to know that I'm not going to die wondering :-).
I've done my research and I've found something that must be the Australian 240V version of the kill-a-watt. I've purchased three and they cost me the equivalent of $US18 each. I've set one up to see what it looks like and am very pleased with how it's performing - after a full 10 minutes of testing before I went home for the day :-). Hopefully when I get past some family duties I'll have more time tomorrow to give it a decent test.
Yesterday, I had performed another upgrade on one of my decommissioned tualatin celerons so that I would have a new machine to test the power meter with. The new hardware consists of
* Rebadged seasonic 300 watt PSU (270 watts rating for 12V rails)
* Asus P5KPL-AM/PS motherboard
* 2GB DDRII-800 generic RAM - single stick
* E3200 Celeron Dual Core (45nm Wolfdale) - stock speed is 12x200 = 2.4GHz
It had a 20GB hard drive with 4GB C:, 4GB D: Windows partitions and the balance was a linux installation with an existing BOINC 6.2.15 install. I did a quick repair install of Windows to get all the new hardware recognised and also installed all needed extra drivers. I will use this Windows later to test the power burning performance of Windows/BOINC to see if there is any difference to what I get for Linux/BOINC. From previous experience, I knew that E3200s seem to run fine at 3.2GHz so, after repairing Windows I decided to set the speed to 12x267 = 3.204GHz. I also decided to leave all BIOS volt options at default settings. I booted Windows and tested the overclock and there were no problems.
After testing the overclock in Windows, I booted straight into Linux where the previous BOINC installation was sitting there ready to go, just as I'd left it a few months ago when the machine (as a PIII) was decommissioned. So I did the usual edit of the old state file to fix things like DCF and the time stats and then setup to get a bunch of APB1 tasks. If I'm going to try to measure credit/watthr or something like that, it would be best to use tasks that have the least possible variability. The machine has been crunching ABP1 tasks overnight and all today so far and quite a few have returned and even validated. There is a little bit of variability in times so far from 15.3Ksecs to about 16.5Ksecs. The average for 12 tasks is 15.878Ksecs or 4.41hrs.
I've now plugged the machine into the new power meter and restarted. I've noted the following power readings
* Plugged in but not powered up - 5 watts standby power
* Booted into the opening BIOS screen - 60 watts steady
* During bootup into Linux - variable from about 50 - 60 watts
* During bootup at the point BOINC started - increase to 83 watts
* Approx 1 min after fully booted - flicking between 79 - 80 watts.
As I said, these figures come from about a 10 minute observation before I departed for the day. The machine is running overnight. The steady reading of say 80 watts with two cores crunching at 100% load seems rather low to me. If the value is correct, that machine is delivering about 2x250/(80x4.41) = 1.417 credits/watthour.
At this stage I've no idea if that's good, bad, or indifferent :-).
The power draw is much lower than expected so the next step is to setup on a second power meter to see how comparable/repeatable the values are.
If the value is correct, that machine is delivering about 2x250/(80x4.41) = 1.417 credits/watthour.
At this stage I've no idea if that's good, bad, or indifferent :-).
The power draw is much lower than expected so the next step is to setup on a second power meter to see how comparable/repeatable the values are.
Trying to support your efforts by providing a couple of bits of data, I found my notes from construction of my most recent machine a Q9550 running stock clock 2.83 GHz with an appreciable undervolt.
The machine to be replaced was a Dell system with a Coppermine near the high end of that part's clock (950 MHz, I think--and thus running with rather severe memory starvation). In October 2008, its Einstein RAC adjusted for resource share (SETI was getting 4%) was just slightly over 100. The power I measured at the plug of the box (i.e. not counting monitor) with Einstein running but the monitor timed out (this mattered because the graphics card chewed an extra 11 watts just keeping the static desktop displayed) was 64 W.
For the new machine I put together, I intended to provide a reasonably capable general-purpose machine, but with pretty high BOINC output, and where feasible made choices for low power. To that end it has a single 1-Terabyte hard drive of the Western Digital line claimed to save some power, and while it has a graphics card it is a fanless model which does not get severely hot to the touch on its modest heat sink.
Parts:
Q9550
Asus P5QL Pro motherboard
floppy drive
Lite-on DVDRW LH-20A1S
WDC WD10EACS 1 Tbyte HD
Agere soft modem
MSI R3450-TD256H Radeon 3450 graphics card
RAM 4Gbyte 4 sticks of DDR2 G.Skill F2-8500CL5-1GBK
Seasonic S12 II SS-380GB power supply
Xigmatek HDT-S1283 CPU cooler
Conditions:
CPU is running stock clock (2.83 GHz) but appreciably undervolted. CPU-Z is reporting 1.080 V at 4xEinstein running. I don't recall the commanded voltage setting, but would hazard a guess that it might be between 1.15 and 1.20 (the quad induces appreciable droop on this mobo when running).
The RAM, so far as I recall, is running dead stock. CPU-Z reports it as 532.8 MHz.
I just moved my niftier Kill-a-Watt from the couple of thousand hours it just spent on my shiny new coffee maker (the way I've been running it was costing me about $7.00/year in power). The nifty thing about the niftier model is, oddly enough, the reset button. You can start an averaging measurement at the moment of your choosing (not last time plugged in, as on the less nifty model). I've just recently pushed the button, so should have a slightly better number tomorrow, but for close comparison with yours:
plugged-in but off 0W after a very brief higher spike charging the supply capacitors
BIOS memtest 95W
Windows User-logon screen while waiting 70W
several minutes after boot, after initial activity died down, running 4 Einstein ABPI flickering from 108 through 110W.
twenty minutes later, after the screen-saver timed out, still 108 to 110W.
Within the time that makes a meaningful contribution to RAC, this machine has been running an almost entirely pure diet of ABP1 reported as the 3.08 application. At this moment it is reporting a 4255 RAC obtained at 96% BOINC resource share, plus a little extra degradation from the fact that my wife uses it somewhat frequently and I occasionally for routine computer work (Solitaire, web browsing, and word processing).
So much for data, herewith a few comments:
Just looking at your host, with a same-generation CPU with half the cores, and with likely lower overhead from non-CPU parts, I think your power measurements may well turn out correct. (I've had four meters of this general class, three models from two makers--all have matched to within a watt or two at the 100W level, and have passed a sanity check using an incandescent light bulb).
My host, while decently efficient, is by no means fully optimized by either component choice or operating conditions for pure crunch cost efficiency. At the component choice level, I suspect I could have done better choosing a motherboard with integrated graphics, and dropping the graphics card, modem, and half the RAM. I also could have used a considerably lower-capacity supply (I chose partly with concern for available plugs, and not sure at the time whether I might go high into overclock and overvolt).
On operating conditions, my undervolt, at 5 BIOS clicks above the lowest setting that gave 1 week of uninterrupted operation, could probably be tweaked lower on a pure crunching host (many report running their overclocks with only one or two clicks of voltage margin).
I did not do a search of speed vs. voltage with any thoroughness (I had real trouble overclocking this rig smoothly compared to earlier experience, and suspect the mobo of having some oddities). Still less did I measure the power across these options, so I have no measurements to back up my guess that I am operating pretty near the power vs. output optimum for my other conditions. (I did make voltage vs. speed graphs of some detail for a previous generation build for which the mobo was better behaved, and am guessing that the general shape of the curve is likely similar here).
With 45 minutes of operating averaging time, and last years incremental power cost, the initial estimate of annual power cost for operating this machine is about $95 in US currency. As I spent at least $1350 for the parts including monitor, a guessed-at present discounted value of likely consumption over the likely life of the machine might be, say $400, so not dominant. But to sneeze at either. For those doing serious economy builds, in jurisdictions with higher power costs, the relative contribution of power would be greatly higher.
( edit ) And Gary has scored two ABP's to date! :-)
I'm quite surprised that Bruce hasn't donned his marketing cap and thought about the potential of this if he were to exploit it a bit.
The pulsar re-discovery page is very nice and I'm sure that those who discover it would be quite chuffed to see their names 'up in lights' for re-discovering a known pulsar. I imagine it's only a matter of time before a previously unknown one gets bagged and that should cause a bit of a stir :-). I read a page somewhere that quoted a figure of around 200K pulsars in our galaxy. With only about 2000 known so far, the potential for new discoveries must be pretty good.
If I were Bruce, I'd offer something like a RAM upgrade kit for every host that participates in a re-discovery. For a completely new discovery, I'd offer a full host makeover to something like an i7 860 - say motherboard, CPU and RAM only, plus a mention in the 'acknowledgements' section of the publication :-). I'd reckon you wouldn't even have to advertise it much - it'd probably quickly become viral all on its own :-).
So instead of crunching less than 37% of the current data gathering rate, with a huge backlog of previously gathered data, the influx of hopefuls looking for a slice of the carrot should be enough to really start making inroads. To get new people joining BOINC projects and, more importantly, to keep them, a relatively small inducement like this should work wonders.
If Bruce was really cunning, he'd approach Intel to donate the prizes in exchange for all the free publicity. In that case you wouldn't restrict yourself to a budget model - how about a full top of the line system. Imagine the photo opportunity for Intel, Bruce, E@H, and the lucky winner as the prize for a new pulsar is handed over.
Bruce, why the hell aren't you doing this???? :-).
RE: I do know they're
)
If it's not an outright red herring that linked in profile (job ad?) has got to be messed up/taken badly out of context. Quite simply there's no such thing as a 128bit integer x86 CPU on the market or being talked up in anyones future architecture slides; nor is there a need for one in anything other than a few specialized tasks. The only explanation that makes any sense is that they'll be making use of 128 bit SSEn instructions; probably in the video pipeline somewhere.
RE: If anyone wants any
)
You've included lots of detail, and are getting admirably high RAC per capital dollar. But I don't spot that you've mentioned power consumption considerations (save only the implications of the power supplies you use).
My impression is that even for people pushing higher toward the bleeding edge in price than you that lifetime power consumption costs are appreciable, and properly considered would cause many to back down to lower overclocks and overvolts than the ones they choose (myself, I choose stock clock and undervolt for this reason).
All other things being equal, your work more toward the value end of the spectrum would be expected to make the relative cost contribution of lifetime amortized power even higher, and thus the optimum overclock overvolt even less.
Is there some reason power matters less to you than it might to some others (someone else paying?...) Even if you don't care for your own circumstance, the (considerable) value of your examples to others would be enhanced with even a little power information. I'd be happy to purchase a power meter and ship it to you if you think you might occasionally use it and report, for example, RAC/watt at the system level. What is local voltage/frequency, and would mechanical socket adapters be acceptable to you?
Great report, Gary. But it
)
Great report, Gary. But it really would be interesting to know about power cosumption of your farm. I would be glad to build something like this - I have enough electricity for only $10 per month :) But all I have now is only a bunch of Celeron's P-III and a lot of pentium mobos. Most of others are not mine - they are in testing after repairments and soon their owners will get them back.
BTW: Gary, what distro do you use in your farm?
RE: RE: If anyone wants
)
Hi Peter, very nice to hear from you.
Yes, you are quite correct to highlight capital dollar. I, most assuredly, should have said, "budget hardware crunching", which is all I was actually trying to talk about.
OK, I did mention the particular Seasonic model and those interested could easily google it. However, having satisfied myself that I would be using a PSU with good efficiency capable of meeting the expected load, I didn't choose to go to the trouble of purchasing a power measuring device simply because I reasoned that there really wouldn't be much I could do to lower power consumption for the conditions I knew I would be using. Let me explain a bit further.
Budget motherboards don't usually come with full voltage control options. If they have those options at all, it's usually for a modest voltage increase. I'm quite aware of the potential power savings from undervolting but the boards I looked at didn't seem to allow it.
What I was trying to do is a bit like planning to drive from LA to New York. First of all I need to choose a vehicle. I could go for a limo or I could choose the cheapest budget Asian vehicle du jour. I reasoned that both would get me there with the level of comfort and efficiency that I desired. So the choice was a no brainer. Next I had to figure out whether I wanted to travel at 70MPH or perhaps more sedately at 45MPH. I could save a lot on fuel (with both vehicles) if I could stick rigidly to 45MPH max. Like most other drivers on the roads these days, I knew that I just wouldn't be able to do that, no matter how much fuel costs I could save.
So I made the decision very early that I was in a hurry to help the project reach its goals and that I wanted to contribute more rather than less even if it meant a higher fuel bill. I regard the level of voltage and frequency increase that I use as being 'relatively modest' compared with what might be the case from using a more expensive board with full voltage options and something better than stock air cooling.
So, in short, electricity costs (though painful) would be something I was prepared to endure.
I'm not sure I understand you here. I assume that the more expensive, more fully optioned boards will consume more power at idle than a minimum featured budget board. I also assume that I'll use less power on the budget board for any given frequency/voltage combo. The more expensive board may be able to win if I was prepared to run at stock frequency with the lowest possible volts. Since I don't wish to run at stock, it's not an option for me.
Are you saying that if I spend $1000 on capital costs and $5000 on lifetime running costs, my relative power costs are higher than if I had spent $2000 on capital costs and $4500 on lifetime running costs? Also, I could do even better by lowering overclock/overvolt so that I got a bit more reduction in running costs? Maybe I could even get the lifetime power costs below $4000 so that I actually saved a bit, overall. If this is what you are saying then I fully agree with you. The only problem is that I'm not prepared to accept the slower pace. I quite like the idea of leaving the voltage at stock and just increasing frequency as much as possible, without compromising stability. I pay the frequency penalty but not a voltage one.
I (or a close family member) pay for every watt used. I'm quite prepared to do that. Let me suggest another analogy. If you increase the price of a pack of cigarettes by an order of magnitude, would everybody stop smoking? Some people are prepared to pay (whatever, almost) for their addiction. I'm prepared to pay for mine :-). I don't smoke, drink or gamble. I'm happy to put those (and other) savings towards electricity. My next project will be to install a mini power station on my roof. We get lots of sunlight in sub-tropical Brisbane :-). If anyone has a viable design, I'm all ears! :-).
OK, thanks for the suggestion. I don't currently own a power meter but I'll look into it. I believe they are readily available locally.
Thank you, most sincerely, for your kind offer but I couldn't possibly accept it. Purchasing one hasn't been high on my list of priorities but it just got bumped (a bit) :-).
Cheers,
Gary.
RE: ... really would be
)
OK, the power meter just got bumped further up the list of priorities ... :-).
Lucky you!! My power bill is a couple of orders of magnitude larger than that :-(.
Not all of that is computers, though ...
PCLinuxOS (based on Mandriva) - mainly 2008 MiniMe but also latest 2009 version.
Cheers,
Gary.
RE: RE: My impression is
)
While businesses often separate capital and operating budgets so thoroughly that they inhibit employees from being sensible, from a pure economics view one would discount the expected future cost stream associated with operating costs over lifetime and add it to the capital cost in comparing alternatives.
In other words, I think of considering the power tradeoffs in cruncher system decisions as just another aspect of the overall quest for best bang for the buck. From a pure cost point of view, for example, this is the sole reason that continuing to run a long-since purchased (and thus free in terms of capital today) antique cruncher is a bad idea, solely because the RAC/watt is so dramatically below modern systems, and that at most jurisdiction's power pricing, that cost is material in crunching.
I'm really not much questioning the wisdom of your choices--most of them look very good to my casual eye. On the other hand as a guide to others, I suspect a stronger inclusion of power cost would probably have made clear to readers that continued operation of your Tualatin fleet was actually well on the way to being big drag on your output per unit cost. Your post to my eye instead emphasizes the opportunity that upgrading them posed to provide much faster harder at a moderate capital expenditure. A balanced view would consider both aspects at once.
Of course, for many of us, it really is not the case of deploying a fixed amount of cash for maximum results--we have floorspace, spousal approval, emotional attachment to machines, and a host of other things pushing us of that optimum. Nothing wrong with any of that--I'm just a fan of rational well-informed choices.
To get back to the single decision of yours that proper consideration of power costs might most likely put in doubt--the overclocking, the possibility would be that consideration of the power cost might lead one to prefer to deploy some of the present discounted value of the future extra cost for power in the form of more machines, or possibly increased capability in some form on the same machines. This might actually be a nonsense in your case, for reasons ranging from rack space to real box availability to a difference in your practical view of present vs. future expenditures.
RE: While businesses often
)
G'day Pete. :-)
First class analysis! Lifetime analysis is often more accurate, at least more revealing. In the computer hardware world capital depreciates so rapidly ( effectively Moore's Law in accounting clothes ), whereas in the power supply world ( energised electrons ) the cost trend is nearly always upwards. So rated against a fixed output ( assuming RAC is some sufficiently and relatively slowly varying 'standard loaf of bread' ) the running costs will always eventually dominate. And probably sooner than you think ... :-)
[ This is a 'what's it worth on a given day' type approach - rather than historical accumulation of how much have has already been committed. This avoids the 'Concorde Fallacy' or as the Irish say "don't throw good money after bad". ]
Cheers, Mike.
( edit ) And Gary has scored two ABP's to date! :-)
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
RE: To get back to the
)
Well, you'll be happy to know that I'm not going to die wondering :-).
I've done my research and I've found something that must be the Australian 240V version of the kill-a-watt. I've purchased three and they cost me the equivalent of $US18 each. I've set one up to see what it looks like and am very pleased with how it's performing - after a full 10 minutes of testing before I went home for the day :-). Hopefully when I get past some family duties I'll have more time tomorrow to give it a decent test.
Yesterday, I had performed another upgrade on one of my decommissioned tualatin celerons so that I would have a new machine to test the power meter with. The new hardware consists of
* Asus P5KPL-AM/PS motherboard
* 2GB DDRII-800 generic RAM - single stick
* E3200 Celeron Dual Core (45nm Wolfdale) - stock speed is 12x200 = 2.4GHz
It had a 20GB hard drive with 4GB C:, 4GB D: Windows partitions and the balance was a linux installation with an existing BOINC 6.2.15 install. I did a quick repair install of Windows to get all the new hardware recognised and also installed all needed extra drivers. I will use this Windows later to test the power burning performance of Windows/BOINC to see if there is any difference to what I get for Linux/BOINC. From previous experience, I knew that E3200s seem to run fine at 3.2GHz so, after repairing Windows I decided to set the speed to 12x267 = 3.204GHz. I also decided to leave all BIOS volt options at default settings. I booted Windows and tested the overclock and there were no problems.
After testing the overclock in Windows, I booted straight into Linux where the previous BOINC installation was sitting there ready to go, just as I'd left it a few months ago when the machine (as a PIII) was decommissioned. So I did the usual edit of the old state file to fix things like DCF and the time stats and then setup to get a bunch of APB1 tasks. If I'm going to try to measure credit/watthr or something like that, it would be best to use tasks that have the least possible variability. The machine has been crunching ABP1 tasks overnight and all today so far and quite a few have returned and even validated. There is a little bit of variability in times so far from 15.3Ksecs to about 16.5Ksecs. The average for 12 tasks is 15.878Ksecs or 4.41hrs.
I've now plugged the machine into the new power meter and restarted. I've noted the following power readings
* Booted into the opening BIOS screen - 60 watts steady
* During bootup into Linux - variable from about 50 - 60 watts
* During bootup at the point BOINC started - increase to 83 watts
* Approx 1 min after fully booted - flicking between 79 - 80 watts.
As I said, these figures come from about a 10 minute observation before I departed for the day. The machine is running overnight. The steady reading of say 80 watts with two cores crunching at 100% load seems rather low to me. If the value is correct, that machine is delivering about 2x250/(80x4.41) = 1.417 credits/watthour.
At this stage I've no idea if that's good, bad, or indifferent :-).
The power draw is much lower than expected so the next step is to setup on a second power meter to see how comparable/repeatable the values are.
Cheers,
Gary.
RE: If the value is
)
Trying to support your efforts by providing a couple of bits of data, I found my notes from construction of my most recent machine a Q9550 running stock clock 2.83 GHz with an appreciable undervolt.
The machine to be replaced was a Dell system with a Coppermine near the high end of that part's clock (950 MHz, I think--and thus running with rather severe memory starvation). In October 2008, its Einstein RAC adjusted for resource share (SETI was getting 4%) was just slightly over 100. The power I measured at the plug of the box (i.e. not counting monitor) with Einstein running but the monitor timed out (this mattered because the graphics card chewed an extra 11 watts just keeping the static desktop displayed) was 64 W.
For the new machine I put together, I intended to provide a reasonably capable general-purpose machine, but with pretty high BOINC output, and where feasible made choices for low power. To that end it has a single 1-Terabyte hard drive of the Western Digital line claimed to save some power, and while it has a graphics card it is a fanless model which does not get severely hot to the touch on its modest heat sink.
Parts:
Q9550
Asus P5QL Pro motherboard
floppy drive
Lite-on DVDRW LH-20A1S
WDC WD10EACS 1 Tbyte HD
Agere soft modem
MSI R3450-TD256H Radeon 3450 graphics card
RAM 4Gbyte 4 sticks of DDR2 G.Skill F2-8500CL5-1GBK
Seasonic S12 II SS-380GB power supply
Xigmatek HDT-S1283 CPU cooler
Conditions:
CPU is running stock clock (2.83 GHz) but appreciably undervolted. CPU-Z is reporting 1.080 V at 4xEinstein running. I don't recall the commanded voltage setting, but would hazard a guess that it might be between 1.15 and 1.20 (the quad induces appreciable droop on this mobo when running).
The RAM, so far as I recall, is running dead stock. CPU-Z reports it as 532.8 MHz.
I just moved my niftier Kill-a-Watt from the couple of thousand hours it just spent on my shiny new coffee maker (the way I've been running it was costing me about $7.00/year in power). The nifty thing about the niftier model is, oddly enough, the reset button. You can start an averaging measurement at the moment of your choosing (not last time plugged in, as on the less nifty model). I've just recently pushed the button, so should have a slightly better number tomorrow, but for close comparison with yours:
plugged-in but off 0W after a very brief higher spike charging the supply capacitors
BIOS memtest 95W
Windows User-logon screen while waiting 70W
several minutes after boot, after initial activity died down, running 4 Einstein ABPI flickering from 108 through 110W.
twenty minutes later, after the screen-saver timed out, still 108 to 110W.
Within the time that makes a meaningful contribution to RAC, this machine has been running an almost entirely pure diet of ABP1 reported as the 3.08 application. At this moment it is reporting a 4255 RAC obtained at 96% BOINC resource share, plus a little extra degradation from the fact that my wife uses it somewhat frequently and I occasionally for routine computer work (Solitaire, web browsing, and word processing).
So much for data, herewith a few comments:
Just looking at your host, with a same-generation CPU with half the cores, and with likely lower overhead from non-CPU parts, I think your power measurements may well turn out correct. (I've had four meters of this general class, three models from two makers--all have matched to within a watt or two at the 100W level, and have passed a sanity check using an incandescent light bulb).
My host, while decently efficient, is by no means fully optimized by either component choice or operating conditions for pure crunch cost efficiency. At the component choice level, I suspect I could have done better choosing a motherboard with integrated graphics, and dropping the graphics card, modem, and half the RAM. I also could have used a considerably lower-capacity supply (I chose partly with concern for available plugs, and not sure at the time whether I might go high into overclock and overvolt).
On operating conditions, my undervolt, at 5 BIOS clicks above the lowest setting that gave 1 week of uninterrupted operation, could probably be tweaked lower on a pure crunching host (many report running their overclocks with only one or two clicks of voltage margin).
I did not do a search of speed vs. voltage with any thoroughness (I had real trouble overclocking this rig smoothly compared to earlier experience, and suspect the mobo of having some oddities). Still less did I measure the power across these options, so I have no measurements to back up my guess that I am operating pretty near the power vs. output optimum for my other conditions. (I did make voltage vs. speed graphs of some detail for a previous generation build for which the mobo was better behaved, and am guessing that the general shape of the curve is likely similar here).
With 45 minutes of operating averaging time, and last years incremental power cost, the initial estimate of annual power cost for operating this machine is about $95 in US currency. As I spent at least $1350 for the parts including monitor, a guessed-at present discounted value of likely consumption over the likely life of the machine might be, say $400, so not dominant. But to sneeze at either. For those doing serious economy builds, in jurisdictions with higher power costs, the relative contribution of power would be greatly higher.
RE: ( edit ) And Gary has
)
I'm quite surprised that Bruce hasn't donned his marketing cap and thought about the potential of this if he were to exploit it a bit.
The pulsar re-discovery page is very nice and I'm sure that those who discover it would be quite chuffed to see their names 'up in lights' for re-discovering a known pulsar. I imagine it's only a matter of time before a previously unknown one gets bagged and that should cause a bit of a stir :-). I read a page somewhere that quoted a figure of around 200K pulsars in our galaxy. With only about 2000 known so far, the potential for new discoveries must be pretty good.
If I were Bruce, I'd offer something like a RAM upgrade kit for every host that participates in a re-discovery. For a completely new discovery, I'd offer a full host makeover to something like an i7 860 - say motherboard, CPU and RAM only, plus a mention in the 'acknowledgements' section of the publication :-). I'd reckon you wouldn't even have to advertise it much - it'd probably quickly become viral all on its own :-).
So instead of crunching less than 37% of the current data gathering rate, with a huge backlog of previously gathered data, the influx of hopefuls looking for a slice of the carrot should be enough to really start making inroads. To get new people joining BOINC projects and, more importantly, to keep them, a relatively small inducement like this should work wonders.
If Bruce was really cunning, he'd approach Intel to donate the prizes in exchange for all the free publicity. In that case you wouldn't restrict yourself to a budget model - how about a full top of the line system. Imagine the photo opportunity for Intel, Bruce, E@H, and the lucky winner as the prize for a new pulsar is handed over.
Bruce, why the hell aren't you doing this???? :-).
Cheers,
Gary.