Green Crunching - Maximizing Credits / Kwatt
8 Mar 2009 1:05:10 UTC
Topic 194223
(moderation:
I would like to hear from other members about how they are building efficient crunching rigs. We have now been into the S5R5 run for over a month now and the credit allocation seems to now be stable.
One rule, this is for E@H credits, not other BOINC project credits to keep this consistent. You will need a Kill-a-Watt type meter for this.
Post key electrical equipment in your rig and describe your typical usage. Measure the Kwh that your rig consumes per day to get credits/Kwatt.
Here are two configurations of mine as examples.
Configuration 1:
CPU: Intel Q9400 2.66 GHz Quad
Mother Board: Asus P5N7A-VM mATX
Memory: 4 x 2 GB DDR2 800
Graphics: Geforce 9300 integrated on MB
OS: ubuntu 8.10, speedstep disabled, no OC
Usage: 24/7 headless server (4 cores @ 100%)
Score: 2645 RAC / 2.75 Kwh = 961
Configuration 2:
CPU: Intel QX9650 3.0 GHz Quad
Mother Board: Asus P5E3 Deluxe (X38)
Memory: 4 x 1 GB DDR3 1333
Graphics: nVidia 8800 GTS 512 add-on card
OS: MicroSoft XP SP3, No OC
Usage: 24/7 desktop (4 cores @ 100%)
Score: 3996 RAC / 4.8 Kwh = 832
It is also interesting to understand the monthy cost to run these configurations. Electric costs $0.10155 / Kwh in my area. Configuration 1 costs $8.38 / month and configuration 2 costs $14.62 / month.
Green Crunching - Maximizing Credits / Kwatt
)
I don't want to be nitpicking but the title should read "maximizing credits / kWh " :-). Anyway that's exactly what you compute in your example as you divide
RAC / (kwh per day) = [credits /day ] / [kWh / day] = credits / kWh
Bikeman
RE: Configuration 1:
)
Wow, 2.75 kWh per day in 24/7 operation means an average power consumption of only 115 Watt. Isn't that a bit low for a Q9400 system? Then again, it's a 45 nm CPU ... It's more efficient than my Merom C2D@2GHz notebook (about 750 credits / kWh). Not bad.
I wonder, however, wether RAC is a good measure because pending credits distort the picture quite a lot.
Anybody with a Core i7 ? Those should really shine in terms of energy efficiency.
CU
Bikeman
RE: I wonder, however,
)
You really need to do a calculation like Krazy Kenzie’s: take the difference between the credit totals at the beginning and end of the measurment period, then adjust for the change in pendings over the same period (remembering that one pending credit is worth 3.35 real credits).
RE: RE: Configuration
)
I think that Roberts values are really to low. Okay i'm crunching with an older Q6700 (Dell E520, 4GB DDR2-667, 7300LE with 22" Acer-LCD, 320GB Sata2 Raid1) and use only 3 Cores for Boinc but my USV tolds me something between 160-168W depends on project. My energy-costs are much higher, i pay 20eu-ct per kwh and boincststs told me an average of ~2300credits per day. I don't know how my USV calculate the value or in which measuring unit possible are apparent or real power. If i calculate my energy consumption through 365 days i get a value from ~11kw per day overall...
I checked my Q8200 which is
)
I checked my Q8200 which is also a 45nm design and with on board graphics and nothing much else but a HD installed, it draws 110 W under full load. I guess the efficiency of the PSU is also important.
CU
Bikeman
I was surprised also by the
)
I was surprised also by the difference between the newer 45nm Q9400 and the 65nm Q6700 at 100% loads. I have both of these in the exact configuration down to the last case screw. The Q9400 ran at a consistent 115 watts and the Q6700 ran at 140 watts, putting the daily power usage for the Q6700 configuration at 3.36 Kwh.
The Q6700 seems to get a slightly better total RAC score of an extra 100 points but the extra power consumed makes it less efficient.
From my second configuration with an add-on graphics card I believe the integrated graphics running headless is what makes these configurations so efficient.
RE: RE: I wonder,
)
Or use Willy's pages to track average daily production. I mean he will give you a 30 day look that you can add up and average the numbers.
I've upgraded configuration 2
)
I've upgraded configuration 2 to the following:
Configuration 2:
CPU: Intel Core i7 920 OC to 3.3 GHz
Mother Board: Asus Rampage ii Gene (X58)
Memory: 3 x 1 GB DDR3 1280 (the default speed for the cpu OC)
Graphics: ATI HD 4670 add-on card
OS: MicroSoft XP SP3
Usage: 24/7 desktop (8 threads @ 100%)
Score: ESTIMATED 5200 RAC / 6.14 Kwh = 847
I had to estimate the RAC, since it will take awhile to reach the full score. Still did not best the efficiency I'm getting from configuration 1. I thought the core i7 would be the most efficient, but even using the score of the best 920 from the "top computers" list of 5623 / 6.14 = 916, this is still short of my efficiency goal of 1000 credits / Kwh.
Interesting to note the difference between 8 and 4 threads is (258w - 218w) 40 watts. So each hyper-thread consumes an additional 10 watts, more than I would have thought.
Monthly cost to run this configuration = $18.71
RE: I thought the core i7
)
This is interesting. I wonder: if we are talking "green crunching", is overclocking the right strategy? Did you have to raise the core voltage of the CPU to get it stable at the increases clock rate? Increasing the voltage by 10 % increases CPU power consumption by roughly 20% (in addition to the increase by the higher clock rate).
Also the external video card is not helping efficiency (unless we are using a GPU app, of course :-) ).
CU
Bikeman
RE: I wonder: if we are
)
So, any businessman or economist would approach this problem in terms of effects at the margin: for a given change in voltage/frequency, does the system level output/watt get better or worse?
One key word in there is "system". If you have a small CPU living in a system loaded with high-power graphics cards, and other power usage not related to computation, you may find that the highest overclock you can get still helps.
On the opposite extreme, a truly green system with Einstein as a major goal would be pared down so much that with a quad CPU, the power directly used in the CPU and indirectly consumed in power supply efficiency is likely to be the considerable majority.
In that case, you are almost certain to find that for at least Conroe and Penryn-class CPUs the system level greenest point is at an undervolt, possibly with the clock rate near or at spec.
I run a Q9550 this way, for which I'll dig up the numbers when I get back home.
I also have, somewhere, a hand-drawn chart of maximum clock speed vs. CPU voltage for my Q6600 system. If I can find it, I may post a picture to help make the point that typically the incremental power efficiency gets drastically worse as you move appreciably above the "comfort zone" for a given sample.