Computer array crunching?

Message boards : Number crunching : Computer array crunching?

To post messages, you must log in.

Previous · 1 · 2

AuthorMessage
Profile ejuel

Send message
Joined: 8 Feb 07
Posts: 78
Credit: 4,447,069
RAC: 0
Message 53232 - Posted: 21 May 2008, 14:28:03 UTC - in response to Message 53231.  

One other thing you should consider, depending on where you live.

Summer is coming up quick in the US. If you live on the Top side of the world, there will be a lot of waste heat generated by an array of computer systems, which will need to be taken care of.

When talking electric bill, just don't forget about cooling all those PCs running at Max.



Right about the cooling...that was in my head the entire time. :) I would have this array in my basement which is actually quite cool in the summer...but of course I will need to watch it. It's a very large basement with nothing in it...and I would only have 20-50 machines so it would represent a very small percentage of space and needed airflow/cooling. Again, all depending and of course depending on if it's 105 degrees every day outside or a mild 80.

As for the overclocking on the previous post, I was/am unaware of how much we could safely overclock. I believe it also depends on the motherboard. If it's safe to get a 30-45% increase, sign me up. I don't think Dell motherboards/BIOS allow me to overclock. Maybe I am wrong.

The whole topic about undervolting/overclocking is over my head. If there is a somewhat technical explanation and how-to, I will read it. But if it's going to be down at the electron level, I don't wanna read that. :) Maybe there is a matrix of simple setups such as the Intel Quad with 5 different motherboards that run just fine.

-Eric
ID: 53232 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile dcdc

Send message
Joined: 3 Nov 05
Posts: 1829
Credit: 115,271,634
RAC: 46,403
Message 53241 - Posted: 21 May 2008, 15:58:01 UTC - in response to Message 53232.  

The whole topic about undervolting/overclocking is over my head. If there is a somewhat technical explanation and how-to, I will read it. But if it's going to be down at the electron level, I don't wanna read that. :) Maybe there is a matrix of simple setups such as the Intel Quad with 5 different motherboards that run just fine.


The adjustments are all straight forward. Actually, there's generally only one thing you need to set to overclock - the FSB speed.

Some useful info:
* Higher voltage allows a faster clock.
* Lower voltage produces less heat.
* The best voltage for any CPU at a given speed is as low as possible.
* Electrical consumption increases with the square of the voltage, and increases linearly with clock speed.

All you do is increase the motherboard's FSB, and the CPU's speed inceases proportionally with it. Run Prime95 (free download) for 6hrs or so to check that it's stable and then you're done. If it's unstable, you can either reduce the overclock, or increase the voltage.

For example, if your CPU is a Q6600 then it runs at 2.4Ghz stock on a 200MHz FSB (often labelled as quad-pumped 800MHz). That means the CPU's mutliplier is 12x:

200MHz * 12 = 2.4GHz

Therefore, whatever you set the FSB at, a Q6600 will run at 12x this speed. If BitSpit's are running at 3.2GHz then he's running them on a 266MHz FSB (quad-pumped = 1066MHz).

You can get more complicated than that but there's no need for Rosetta. Just make sure the motherboard you're buying supports 1066MHz and allows you to OC and you're good to go. There'll be plenty of people to suggest good motherboards for that here.
ID: 53241 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
BitSpit
Avatar

Send message
Joined: 5 Nov 05
Posts: 33
Credit: 4,147,344
RAC: 0
Message 53258 - Posted: 22 May 2008, 0:24:19 UTC

I ran 16 machines in the basement here at one point. Trust me. The temperature will go over 80. Computers double as decent space heaters.

And to correct a few of dcdc's numbers, and Q6600 has a multiplier of 9 so...

Stock Q6600, 2.4GHz, 266 FSB, DDR2 553 minimum
Q6600 @ 3.0GHz, 333 FSB, DDR2 667 minimum
Q6600 @ 3.2GHz, 356 FSB, DDR2 800 or quality 667 with some slightly loosened timing
Q6600 @ 3.6GHz, 400 FSB, DDR2 800
ID: 53258 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile ejuel

Send message
Joined: 8 Feb 07
Posts: 78
Credit: 4,447,069
RAC: 0
Message 53260 - Posted: 22 May 2008, 2:24:27 UTC - in response to Message 53258.  

I ran 16 machines in the basement here at one point. Trust me. The temperature will go over 80. Computers double as decent space heaters.




Oh I hear ya. :) I just mean I have a REALLY big basement so I would hope that 20+ pcs wouldn't affect the temp too much down there before needing some cooling. :)
ID: 53260 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Mod.Sense
Volunteer moderator

Send message
Joined: 22 Aug 06
Posts: 4018
Credit: 0
RAC: 0
Message 53276 - Posted: 22 May 2008, 14:54:23 UTC
Last modified: 22 May 2008, 14:54:45 UTC

Calculate your power requirements. If each unit takes 200 watts, it uses that for calculating... but then all basically becomes heat. ( reference ) So picture a 4000 watt space heater down there... and keep in mind that a space heater has a thermostat that turns it off for periods of time; but your machines will be putting out heat full time. And you also must picture at least 3 15amp circuits to your basement, and 100 kwh per day electric bill (>$300/month?).

Not trying to discourage you. Just trying to help you plan accordingly.
Rosetta Moderator: Mod.Sense
ID: 53276 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Paul

Send message
Joined: 29 Oct 05
Posts: 193
Credit: 65,742,447
RAC: 984
Message 53295 - Posted: 23 May 2008, 9:59:50 UTC - in response to Message 53276.  

Power and cooling are my two major issues right now. I have 11 systems running and 7 of them are stacked in 1 room in my basement. Normally my basement is cool but I can not get the temp below 75 in that room. I watch the temp because I don't want to kill my PCs.

I found that you can often get private universities and other organizations to give you older PCs when you explain the R@H project. I have several systems that were donated. All of them are P4s and most of them support HT. I pack them full of inexpensive DDR RAM and watch them run. I also watch ebay and the Dell 530 (quad core with 2 GB RAM) are now below $400.

We installed a geothermal HVAC system and this will be my first summer with all 11 systems running. My electric bill is usually about $400 month so I can't wait to see how much higher it will go. The P4s are very power hungry and that means lots of heat for every calculation.

I would really like to find a way to just stack MBs with a shared PSU and netboot or USB boot the whole stack. I have a keyboard/video/mouse switch with 5 or 6 units connected.

I would like to add a couple more quad cores to my collection. They overclock well and use very little power compared to some of the older CPUs.

Thx!

Paul

ID: 53295 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Jack Shaftoe
Avatar

Send message
Joined: 30 Apr 06
Posts: 115
Credit: 1,307,916
RAC: 0
Message 53311 - Posted: 24 May 2008, 14:07:16 UTC - in response to Message 53241.  

All you do is increase the motherboard's FSB, and the CPU's speed inceases proportionally with it. Run Prime95 (free download) for 6hrs or so to check that it's stable and then you're done. If it's unstable, you can either reduce the overclock, or increase the voltage.


I have always wondered: why use Prime95 instead of BOINC? Does it give you lots of info about stability that a failed WU or a system crash won't?

I've been itching to OC as well, but each time I try I get blue-screen or system freezes which I can't figure out how to resolve so I go back to stock.
ID: 53311 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile dcdc

Send message
Joined: 3 Nov 05
Posts: 1829
Credit: 115,271,634
RAC: 46,403
Message 53313 - Posted: 24 May 2008, 15:15:28 UTC - in response to Message 53311.  

All you do is increase the motherboard's FSB, and the CPU's speed inceases proportionally with it. Run Prime95 (free download) for 6hrs or so to check that it's stable and then you're done. If it's unstable, you can either reduce the overclock, or increase the voltage.


I have always wondered: why use Prime95 instead of BOINC? Does it give you lots of info about stability that a failed WU or a system crash won't?

P95 increases the load on the system quite a lot more than R@H, but more importantly P95 compares the results that your computer produces against known values, whereas R@H doesn't have the known true values to compare against.


I've been itching to OC as well, but each time I try I get blue-screen or system freezes which I can't figure out how to resolve so I go back to stock.


What settings are you trying and on what machine?
ID: 53313 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
BitSpit
Avatar

Send message
Joined: 5 Nov 05
Posts: 33
Credit: 4,147,344
RAC: 0
Message 53356 - Posted: 26 May 2008, 17:22:05 UTC

Okay, I finally got around to whipping up a page for my setup. It's at http://splicedcollective.org/armada/ It has my SMP kernel, control script, directory structure layout, and some pictures of my stuff.
ID: 53356 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
glaesum

Send message
Joined: 16 Oct 06
Posts: 21
Credit: 508,632
RAC: 1
Message 53361 - Posted: 27 May 2008, 0:24:33 UTC

interesting reading from BitSpit et.al. though I'm not really techie enough to understand it all.

there has been an open source community developing a firmware mash-up to turn the hundred dollar Linkys NSLU2 ["Slug"] into a unrestricted Linux server. it has an ARM processor inside but not much memory. I wonder if it has any use {for the very hardcore!} in the crunching arrays being discussed above - perhaps only as a low cost/low power server even if not as a main cruncher...

http://www.nslu2-linux.org/wiki/Main/HomePage
http://en.wikipedia.org/wiki/NSLU2

(people have managed to use it as email/web/music server, home PABX phone switch using Asterix, etc.etc.)
ID: 53361 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Jack Shaftoe
Avatar

Send message
Joined: 30 Apr 06
Posts: 115
Credit: 1,307,916
RAC: 0
Message 53377 - Posted: 27 May 2008, 11:32:08 UTC - in response to Message 53313.  
Last modified: 27 May 2008, 11:33:11 UTC

What settings are you trying and on what machine?


GA-EX38-DQ6, Q6600, 2GB OCZ DDR2 800. I tried changing the FSB to 333 MHz, which should give me around 3GHz, but it won't boot.

If I can figure that one out, I also have a GA-G33M-S2L board with a Q6700.
ID: 53377 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile dcdc

Send message
Joined: 3 Nov 05
Posts: 1829
Credit: 115,271,634
RAC: 46,403
Message 53378 - Posted: 27 May 2008, 11:49:12 UTC - in response to Message 53377.  

What settings are you trying and on what machine?


GA-EX38-DQ6, Q6600, 2GB OCZ DDR2 800. I tried changing the FSB to 333 MHz, which should give me around 3GHz, but it won't boot.

If I can figure that one out, I also have a GA-G33M-S2L board with a Q6700.

did you leave the memory at 200MHz (aka 800MHz) or was it linked (and therefore also overclocked)? You want to keep the memory at stock and just up the FSB/CPU.

My Q6600 won't boot at 3GHz at stock voltage either, although my Q6420 will do 3.2GHz at stock... I've not had a chance to play with the Q6600 yet though.
ID: 53378 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Paul

Send message
Joined: 29 Oct 05
Posts: 193
Credit: 65,742,447
RAC: 984
Message 53380 - Posted: 27 May 2008, 13:48:33 UTC - in response to Message 53377.  

What stepping of the Q6600? B3 or G0? On my B3, I really have to over voltage to get it going. On my G0, it requires far less.

You need not make the jump directly to 333. Start with 269, 279, 289 and 299 and unlink your memory. I have one unit with some high priced 800MHz RAM that just will not overclock much. I had to uncouple the RAM and CPU to get things going.

You will find lots of information on overclocking on several websites but don't feel like you have to make 1 jump to 3GHz. Start small and document your bios configs. You will find it is fun stuff to push those chips a little harder.

good luck!

Thx!

Paul

ID: 53380 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile dcdc

Send message
Joined: 3 Nov 05
Posts: 1829
Credit: 115,271,634
RAC: 46,403
Message 53385 - Posted: 27 May 2008, 14:12:02 UTC - in response to Message 53380.  

What stepping of the Q6600?

mine is G0 - currently undervolted to 1.1V . I'll have to have another go on the OC at some point.
ID: 53385 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Ingleside

Send message
Joined: 25 Sep 05
Posts: 107
Credit: 1,514,472
RAC: 0
Message 53398 - Posted: 27 May 2008, 18:59:50 UTC - in response to Message 53311.  
Last modified: 27 May 2008, 19:04:42 UTC

I have always wondered: why use Prime95 instead of BOINC? Does it give you lots of info about stability that a failed WU or a system crash won't?

I've been itching to OC as well, but each time I try I get blue-screen or system freezes which I can't figure out how to resolve so I go back to stock.

Using BOINC as a test is a possibility, but you must either re-run the exact same wu's and compare the results afterwards, or run a project that uses a quorum.

Just running Rosetta@home won't be a good indication, since even if it hasn't crashed, you can still have generated an invalid result. Also, even Rosetta@home validates all results, even if you've got an "impossible" result that aren't used for scientific purposes, you'll AFAIK still get your credit...

Prime95 is an "easy" test, since if there is any calculation-errors, it will immediately be detected.

But, Prime95 and various DC-projects does not use the cpu and memory the exact same way, so even if you've done a "stable" Prime95-run, it does not mean running a DC-project will not use the "wrong" part of the system and starts generating garbage...

Overclocking for so running 24-hour-tests of Prime95, Memtest, Gromacs and so on error-free means "no problems detected yet". It does not mean "computer won't make a wrong calculation due to overclocking once a month"...
"I make so many mistakes. But then just think of all the mistakes I don't make, although I might."
ID: 53398 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Jack Shaftoe
Avatar

Send message
Joined: 30 Apr 06
Posts: 115
Credit: 1,307,916
RAC: 0
Message 53440 - Posted: 29 May 2008, 16:40:02 UTC - in response to Message 53380.  
Last modified: 29 May 2008, 16:41:03 UTC

What stepping of the Q6600? B3 or G0? On my B3, I really have to over voltage to get it going. On my G0, it requires far less.

You need not make the jump directly to 333. Start with 269, 279, 289 and 299 and unlink your memory. I have one unit with some high priced 800MHz RAM that just will not overclock much. I had to uncouple the RAM and CPU to get things going.

You will find lots of information on overclocking on several websites but don't feel like you have to make 1 jump to 3GHz. Start small and document your bios configs. You will find it is fun stuff to push those chips a little harder.

good luck!


How do I unlink my memory? Not sure I understand this. G0 Q6600 is what I am working with. I am sure unlinking is different for each MB, but what do you typically see this setting described as?

Thanks Paul,
ID: 53440 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Paul

Send message
Joined: 29 Oct 05
Posts: 193
Credit: 65,742,447
RAC: 984
Message 53543 - Posted: 5 Jun 2008, 10:38:39 UTC - in response to Message 53440.  

To unlink the memory, look for anything that lists the speed of the RAM. Usually you will see that it is locked at 1.5x the FSB and this yields 800MHz (or 667 depending on your RAM). You can usually change that multiplier and you want to because you can over run your ram. Initial overclock settings of 1:1 or 1.2:1.

I don't remember exactly where that setting is on my MSI boards but now that you know what to look for it should help.

If you still have trouble, send me a PM. We can figure it out.
Thx!

Paul

ID: 53543 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Previous · 1 · 2

Message boards : Number crunching : Computer array crunching?



©2024 University of Washington
https://www.bakerlab.org