A new farmer in town...

GameALot

Weaksauce
Joined
Mar 13, 2005
Messages
67
Hi, All!

I have been lurking in these forums for some time, basking in the greatness of all things [H]. While I am a new poster, I have been folding for several years now, just not full time. However, a couple of events caught my attention.

First, there was the link to the mini-itx cluster. Aye, she was a sight! But, I couldn't see an easy way to get one to fold.
Then came the post by unhappy_mage regarding his Fold-Server distro. Damn, I was hooked!

Much research ensued. I read the spreadsheets by OSUguy98. Delved into the opinions in the processor and motherboard forums. And I came up with a plan. I just wanted to run it by you all before I committed.

I intend to use one of the new 90nm Sempron 3100+'s on a DFI K8M800-MLVF, as suggested by OSUguy98, with 512MB RAM for the server, with similar set-ups for the clients. I am hoping to get the best trade-off between starting $,kWh,and speed. My only question is how this Sempron compares with the Athlon XP's.

My plan is to build the server and see how much it's going to cost to run. Then add clents as kWh's allow.

I will now open the floor to all. Forgive me if this has been covered before, Google is my friend, he just didn't help much. And my head is starting to hurt.

Before I go, I just want to say thank-you to everyone here. I hope to add more than boxen to the [H]orde.

 
First, Welcome!

Others are better with the numbers on $/Point and all that.

My only advice is take a good long look at the base memory your get for the clients Pick the types of WU;s you want to run, and let that determine the memory expense.

Memory requirements for the big WU's is on the up and up. 600 pointer normal big WU's honestly don't take as much memory as you think. Basing the clients as 100% folders, no human use, running 256 meg ram is not out of the question. I'll be confirming this shortly, but they (600 pointer normal WU's) only take ~104 meg ram. Tack on the net book Linux client, I'm to try and do this with 192 meg of ram. This is on older units, no reason in the world to buy a 64 meg stick for you so 256 is a nice round number.

However, QMD cores are insane. I've got some which take ~317 meg ram by themselves. I actually thing I've got some newer 600 pointer QMD cores, need to check these out more.

I'll (we'll) be posting much more about unhappy_mage's net boot fold server very soon. In a word, it works. Way cool. He is also willing and eager to improve it (speaking for him).

From a point's point of view, big WU's score major bonuses. However, who know how long that will last. QMD cores have a greater bonus from what I can tell than regular big WU's.

From reading over at Stanford, it's kinda looking like overall memory requirments are on the rise. I'd feel funny about spending the money on 512 meg just for a folding box, that's just me. However, it might be needed- depending on what types you want to run. Can always buy 256 now and add later. Keep the memory in dual channel mode, folding is big time dependent on memory bus from what I can tell.

Fake edit: your only question was about CPU types. Opps, I'm posting this anyways since I typed it all up.
 
Hey, Marty! I was hoping to hear from you.

These will be dedicated folding boxes. The server will definately have 512MB. I may start the clients with 256MB, I just don't want to limit my expansion possibilities.

It has been tough getting info about the 90nm Sempron 3100+. It seems that they aren't officially released yet. ;)

I chose the m-atx format in the hope of reducing the power costs over time. I'm not too concerned with overclocking the board. But a little tweaking MUST be done.

You mention running the memory in dual-channel mode. I have read that none of the socket 754 boards will be able to use that. Something about a pin missing.

Thanks for the quick reply. I look forward to hearing more about the fold-server software.
 
From what I've seen with my little windowbox garden, If you've got a GH+ of RAM the big pointers are coming you're way. It's not really needed, but the supercompter at Stanford, figures you can handle it. The 2 most powerful, of my modest 4 rigs, have run nothing but 125 point QMD's or higher. For the last 12 weeks.

 
Here's a data point:

My Sempron 3000+ @ 2007MHZ (S754 90mm) and Athlon XP-M 2500+ @ 2300MHZ (Barton) are both chewing on a P1141, Gromacs 600 pointer, and are running neck and neck @ around 35m30s or 10.16 PPH, 243.84PPD, 1706.88PPW

Hope that helps.
 
:D Glad that you found the spreadsheet I threw together useful!! (Know of any way I can improve it??)

Good luck on the endevour of building a farm... It's been a dream of mine for about a year now, but I don't have the funding, nor the space to do it... If you decide to go case-less, and need a little help drafting up a solution for your blades, let me know... I'm pretty good with AutoCAD and/or Microstation...

As far as I know, the Sempron 3100+ is based off of the A64 chips, while all the other Semprons (so far) are based off of the Athlon XP chips... So it seems to me that whatever gains you have going from an AXP to a A64 would be a good base point for comparison with the Sempron 3100+... (Though I've been known to be wrong!)

Let us know if you have any more questions,etc... we're all here to help!


Keep on Folding!!

 
i find myself itching to build myself a farm.....

but the money think is very hard to come by, maybe i can stop drinking..... or n ot... i drink and try and borg more


 
while i cant help you on the cpu question I would like to say Welcome to the team!!

Buck
 
I kinda mis spoke on the RAM in dual channel. It would be pretty hard to run 256 in dual channel, unless you bought 128 meg sticks. Buying a 128 stick in this day and age would be a little silly in my opinion. CPU type will matter more in the long run than memory config I'm betting.

I'm sure AMD chips would be the way to go. Folding is 90% CPU clocks though, and some big WU cores not liking AMD's SSE2 stuff. (I'm not an AMD chip expert if you can not tell....)

I priced out a Celeron D set (MB/CPU/RAM/PSU) and for ~$200-225 you could get a naked 3.06 GZ folder. These things overclock easy (buy the ones with a higher voltage core, E0 stepping voltage 1.2?? to 1.400 voltage listed on newegg). Many people get 4 GZH on stock air from these guys. 4 gzh nodes is attractive, offical SSE support (I'm sure AMD's SSE2 will be supported soon). 533 FSB means the ram FSB would not get too high on the OC, PC3200 should give you some room. Price per KW used might be a different story.

Someplace on the net, it is possible to run two motherboards from one PSU. This would keep that constant price down. Some guy ran a stack from one big AC/DC converter skiping the PSU all together.

I'm not trying to start a AMD/Intel flame thread, I could care less. I think the best producers are the LV Xeons (some guy had a dual 1.6 Xeon OC'd to 3.something---dang). Those Xeons are spendy as heck.

Drop ChelseaOilman a PM. He went through all this, ask him what he did, and what he would do differently had he to do it all over again today. He knows what he is doing also, or does a great job faking it.... :)
 
Not a bad setup:

Sempron 3000+: $96 (ZipZoomFly)
DFI K8M800-MLVF: $73 (ZipZoomFly)
512MB: $46.50 (PQI Power Series @ Newegg)
PSU: $24.50 (Enermax 300W EG301P-VB-1-FM)

Price Point: $240 (not bad for a low-voltage 160-280 PPD machine)

I like the 3000+ because you don't need the extra 128KB Level2 cache of the 3100+. They have the same 1.8GHz speed.

I really don't think you would need a PSU above 300W if you are folding. You could probably even go lower, but....(the more loaded a PSU is the more efficient it is)

I have heard of pretty good success undervolting the 90mm AMD chips. If the DFI board supports this, it should drop your power usage also.

edit: Marty, no dual channel on the S754s


 
Don't we have a link or something in the FAQs for splitting a PSU? I'm pretty sure Turmelle did this, maybe one of his threads had a step-by-step with pics... Though, that thread may have been lost in the forum purge last year...


Keep on Folding!!

 
My God.

You guys type to much.

And this topic is WAY on topic. no fun.

But welcome ;) and happy folding!

and remeber, send all your quarters to me. I have to pay for the wash machines
again. and i have a real job. so i need to wear real clothes, that are clean. not dirty.

pwned!!
 
I'd have to agree with marty that the Intel route is unfortunately the better way to go at the moment.
From Stan4rd:
# What about SIMD (SSE/SSE2/3D-Now) support? QMD core uses double precision only. Therefore, only SSE2 will be useful. Currently, Intel Pentium 4 SSE2 is supported, and AMD Athlon64 SSE2 will be added soon.
 
Come to the dark side my young one...

KodiakStar- you need to get laid.

I have not had a post deleted this week, now's a good time to start..

Fake edit: so do I
 
GameALot said:
These will be dedicated folding boxes. The server will definately have 512MB. I may start the clients with 256MB, I just don't want to limit my expansion possibilities.
...
Thanks for the quick reply. I look forward to hearing more about the fold-server software.
The server doesn't need any more memory than the clients. Right now, looking at 'top', I see that I've got 100 MB of memory used, of which 60 is cached data (meaning not in use by a program), so the real total is about 40 MB. It's not running a WU right now, but like marty said: 104 for bigpackets gromacs, 350 for QMD, and something like 25 for normal packets. I think timeless WUs are only around 5-10, but you probably don't need to worry about that.

One thing to note, though: If fah accidentally sends you a packet that doesn't fit in memory on fold-server's clients, then it will immediately curl up and die. 256 will be enough if you're not planning to run QMD, and that's what I'd get initially, then when it's time to upgrade, get 1 512 per 2 machines and put the 2 256s together in one.

Fold-server is designed to be easy to set up and keep going, and I can definitely help you with the process if Linux isn't your bag. The docs are helpful, give them a read before you get too deep into it. You can PM me or email me. I'll be glad to help you out getting your farm going.

Fold on, bro!
 
unhappy_mage said:
One thing to note, though: If fah accidentally sends you a packet that doesn't fit in memory on fold-server's clients, then it will immediately curl up and die. 256 will be enough if you're not planning to run QMD, and that's what I'd get initially, then when it's time to upgrade, get 1 512 per 2 machines and put the 2 256s together in one.

Umm, little late for that info..... geezzzzzzzzzzzzzzzzzz I'm kidding, mostly.

QMD cores/ Allowing big work units gerater the 5 mb- This is important, is Stanford doesn not allow a -qmd flag (Allow download of big WU's, and NOT QMD cores or vise verse, another way to put this if is the standard allow big WU's means you automatically can get QMD cores this will really screw things up for alot of people).

There is talk of needing to have a -qmd flag, guys running HT chips/Xeon with more than one instance allowing big blocks are getting killed by the memory usage. 317 meg X 4 on a dual Xeon board, yikes. 1.2 Gig used for folding, and your also running a server OS for many of these guys, their pissed right now.

Mage- curl up and die and start over, or die as in dead stick pull plug from wall!
 
gnewbury said:
I'd have to agree with marty that the Intel route is unfortunately the better way to go at the moment.
From Stan4rd:


Yeah, my 1.4Ghz Opteron is outproducing my AthlonXPs on certain work units and I only expect that to become more apparent.

Stanford is also saying that SSE3 is not important for what they are doing, so Sempron 3000+ should be fine (have SSE2).

In other words, XPs are on the way out, but any 754,939,940 solution should still be very competitive against Intel.
 
From what I understood, you only get the qmd units if you have the -advmethods flag AND big packets turned on... (as said above, I've been known to be wrong!)

As stated above, you don't really need more RAM in the "server"...

My advice for RAM, if you can swing it, get the 512MB of RAM... WU are only getting bigger... Or... get 256MB, and have more boxen....:D


Keep on Folding!!

 
marty9876 said:
Mage- curl up and die and start over, or die as in dead stick pull plug from wall!
As in crash and reboot without doing it the nice way. This means you'l probably get a new unit, since it won't get saved to the server, but it means Stanford will have given out a unit that'll never get finished, so they have to wait until the deadline. It's not a huge problem from your perspective, but Stanford might get a little upset if it happened a lot.
 
GameALot said:
You mention running the memory in dual-channel mode. I have read that none of the socket 754 boards will be able to use that. Something about a pin missing.

Yeah about 185 pins.

Welcome!
 
Jeez... I get a little sleep, and look what happens! I knew I came to the right place.

@neisius: I looked into the dual Zeon set-up, but decided to stay with single processors. It will be easier to break up the farm and sell complete systems if I decide to upgrade, or I really need the money.

@Hito: The 3000+ looks like a good idea. Would the extra cache on the 3100+ be more future proof? It's less than a $20 difference at NewEgg.

@OSUguy: Yes, these boxen will be nekkid, headless, and diskless. Thanks for the offer of drafting assistance, I will look you up if I have any trouble. My research leads me to believe that the Palermo core Semprons are mobile A64's with no 64bit and partially disabled cache.

@Marty: I have concentrated my efforts on an AMD solution. But, I will take your advice and look into the Celeron's. BTW, are you talking about the Prescott core ones? I agree that the Zeons look good. But yes, they are very $$$.

@unhappy_mage: I will follow your suggestion regarding the RAM. I have been reading your docs. and browsing the DSL pages. I have some questions, and will e-mail them to you today if you don't mind.

I read the articles about splitting the power supply, but I would rather have one for each motherboard. That way, only one needs to be down at a time for maintenence,etc. Plus, if the supply goes, it will do less damage. I think I'll run with Seasonic supplies, they are recommended here in the forums, and have a good price point. I will be running a UPS too, just to be safe.

When do you think AMD SSE2 will be supported? I wonder how that will change the choice between Intel and AMD. I would start with AMD if SSE2 support is coming relatively soon, and will make a difference.

Back to power... My main machine is a 3ghz P4 with 1gig of RAM, a Ati x700-pro,2 sata HD's, DVD burner, and misc. case fans running off of a 485W CoolerGiant PSU. My amp meter shows 1.5amps usage while running FAH, which comes to 180W if I'm correct. Do you think I could use 200W PSU's in the farm. The Seasonics seem to be understated when it comes to their capabitilty.

Damn, that was a long post. :rolleyes: I hope I didn't lose anyone. The good news is that I'm done for now. :)

Thanks for all the support.


 
GameALot said:
@unhappy_mage: I will follow your suggestion regarding the RAM. I have been reading your docs. and browsing the DSL pages. I have some questions, and will e-mail them to you today if you don't mind.
...
When do you think AMD SSE2 will be supported? I wonder how that will change the choice between Intel and AMD. I would start with AMD if SSE2 support is coming relatively soon, and will make a difference.

Back to power... My main machine is a 3ghz P4 with 1gig of RAM, a Ati x700-pro,2 sata HD's, DVD burner, and misc. case fans running off of a 485W CoolerGiant PSU. My amp meter shows 1.5amps usage while running FAH, which comes to 180W if I'm correct. Do you think I could use 200W PSU's in the farm. The Seasonics seem to be understated when it comes to their capabitilty.
Email will work fine. Hope I can answer all you got!

Never. I don't think they're going to bother with it.

I have a 200w psu powering a 450mHz p3, and 4 hard drives. 200w should be plenty. the psu calculator may say otherwise, I haven't tried.
 
GameALot said:
@Hito: The 3000+ looks like a good idea. Would the extra cache on the 3100+ be more future proof? It's less than a $20 difference at NewEgg.


At the price point we are talking about, $20 is a 20% price increase for, as far as I can tell, no increase in performance for folding. Based on that, I guess I would say it is not worth it. Spend the money on a better HSF or get a free box after you buy 12...... ;)
 
Cache has no/little impact from what I've read at the offical F@H forum. My understanding is the memory "blocks" are too big to fit in a CPU cache, hence my thoughts on running dual channel. The folding program needs to referance large blocks (larger than the cache), so that's my take on it. I could be wrong there. Not saying it has no increase, just way less then one might think.

All my thinking is on max points per upfront dollar spent upfront, there are reasons for this. Take that for what it's worth (not the correct thinking for someone running this at home...ie paying the power bill).

I'd like to see an overclocked P4 Celeron 3.06 @ 4.0 on OSUguys spreadsheet. I'd be curious to see where it fits in in terms of point output and such.

SSE2 on AMD- I would not count on anything from Stanford, I'm sure they have alot going on over there. One never knows, but the market shares of AMD to Intel are still 90% Intel so I'm sure this is where their effort is. It's just serving the masses here, no AMD/Intel bashing going on here folks...move along now.... :)

I'm 99.9% sure AMD's are the way to go regarding points/upfront/long term power usage. I only suggest you look at the Celerons because of their increased output (maybe?) while running big blocks.

My memory thoughts are I see a lot of "old timers" running normal WU's it seems on their farms and I'm just not sure why. It's all for a good cause, don't get me wrong. From a points point of view I just wonder if they might be able to put out more.

My CPU count is way, way below others and I'm able to keep up with them at times. No pesky humans on my boxen helps big time. Also it's not apples to apples comparing copr borging to setting up a personal farm.
 
I have a 2.53Ghz Celeron 1066FSB running. I will get off my butt and try to get some statistics on it. If you are paying the bill though, Low voltage AMD is going to be the answer. If upfront cost isn't a big issue, I like the Pentium-M solution, but you are talking around $400+ per box. They run at lower voltage though (to my knowledge) They are the best processor for Big Packet Gromacs, by a long shot. But, you could probably buy at least 3 Sempron box for the price of 2 Pentium M boxes and still have money left over.

My Pentium M 1.8Ghz does a P1140 Step in 31 minutes (versus Sempron 3000+ @ 2Ghz completion in 36 minutes) which is almost a 35PPD advantage or 12,700 PPY advantage.
 
marty9876 said:
My memory thoughts are I see a lot of "old timers" running normal WU's it seems on their farms and I'm just not sure why. It's all for a good cause, don't get me wrong. From a points point of view I just wonder if they might be able to put out more.

My CPU count is way, way below others and I'm able to keep up with them at times. No pesky humans on my boxen helps big time. Also it's not apples to apples comparing copr borging to setting up a personal farm.

Marty,

I think there are 2 reasons: 1. "Old timers" borged or invested in older boxes so have more sub 2GHz boxes running; 2. Many people have a lot of computers borged. Between the memory usage of big packets and the case that non-dedicated boxes usually aren't running 24/7, they are much more in-efficient.
 
Hito- Non dedicated boxes I totally understand not running big WU's, that's just wrong to on those borgs.

As far as sub 2 GZH, I've got a 1.2 that does just fine on big WU's. I think she does 120 PPD on big WU's, and 53-80 on normals. I don't think GZH and big WU's really matter, granted it does take some time to turn them, but the PPD seems to be worth it.

When I say "Old Timers", I really mean people whom have been doing this far, far longer than I and I'm just trying to learn form them, understand what they are doing. I'm thinking they know more than I do.
 
Marty... How about we collect some data? In another thread, we could gather up as much "reliable" info about alot of different CPUs and their ppd... (I only say "reliable" because the more people you have throwing data into the mix, the more deviation there will be)...

I can throw together another spreadsheet with stock CPU, actually clock, OCed clock,etc... and then we can figure out how best we can relate all this info....

I'm up for trying... but we'd need a LOT of info... I have the last year's worth of info on my processors... I just haven't sat down to figure out how to relate all the data....


Keep on Folding!!

 
marty9876 said:
Hito- Non dedicated boxes I totally understand not running big WU's, that's just wrong to on those borgs.

As far as sub 2 GZH, I've got a 1.2 that does just fine on big WU's. I think she does 120 PPD on big WU's, and 53-80 on normals. I don't think GZH and big WU's really matter, granted it does take some time to turn them, but the PPD seems to be worth it.

When I say "Old Timers", I really mean people whom have been doing this far, far longer than I and I'm just trying to learn form them, understand what they are doing. I'm thinking they know more than I do.

Marty, what I meant by Sub 2Ghz boxes is that old timers invested in older equipment that doesn't produce as many PPD as newer boxes. In my case, all my gear (excepting one unfortunately Lappy) is running at 2100+ speeds, but I know a lot of the CPUs out there are 1GHZ Pentiums and older.
 
Hito Bahadur said:
Marty, what I meant by Sub 2Ghz boxes is that old timers invested in older equipment that doesn't produce as many PPD as newer boxes. In my case, all my gear (excepting one unfortunately Lappy) is running at 2100+ speeds, but I know a lot of the CPUs out there are 1GHZ Pentiums and older.

Not an old timer but most of my boxes are sub 750mhz.
 
OSU- Not to cross ships here, but they are doing that over at OCAU right now. I just have not checked it out yet. Yea, yea I'm so folding freaky I read all the forums.....

I've got tons of data, but really they are only across two types ot boxes. I've got others, just not many of them. Plus all the crappers just light up..... Ok maybe I do have data.

No data on a Celeron OC'd to 4.0, since I'd have to buy that (I might, but I want to OC it to 5 GZH, which I might not be able to so it ruins the whole mess for me)

This is still on topic of some thread I forget the title of right?

Great ideas all!
 
Spectre said:
Not an old timer but most of my boxes are sub 750mhz.


Spectre, I understand. I was answering marty's question about why he blows through "old Timers" when he has many less CPUs. he blows past many non-old timers also (me included).
 
One this is for sure I think, before big WU's building a farm was different than after the release of big WU's, correct?

Let me know if I'm off here.
 
I think you're right... now, RAM is brought into question... But I don't think it was the big WUs that brought out the red flag for RAM... I think it's the new cores, and the ever-multiplying number of new cores, that brought up the question... with the timeless tinkers, gormacs and regular tinkers, RAM wasn't really an issue... but double gromac core, the amber core, the newest QMD core have all increased the amount of RAM used...

Stanford realizes that many are just using their home PCs for this, but you're basic minimum (read: new) Dell is going to be around 2.4GHz... As PCs get faster, and RAM cheaper, the WUs will just get bigger... and bigger


Keep on Folding!!

 
Back
Top