Just Switched my Folding (FAH) to TechIMO, i7 on Vista, GPU questions

[Sky King]

Hello all,

I have been folding for a few years, but always on a spare CPU here or there, so I have only garnered about 30,000 points in about 2 years. Well, that has changed dramatically. I have just up'ed my output to a little over 2,000 PPD and am on my way to doubling that over the next few weeks, so I needed to find a team where I might be able to get some occasional tech support for more advanced folding, such as SMP and GPU. A quick visit to the TechIMO forums, and seeing how active the group is, I knew "this is it, baby"... So I switched over a few days ago (User xam.relseor) and have already pumped in something like 6,000 points since last weekend. My diabolical goal is to get to around 4,000 PPD, so that I am putting 120,000 points per month into TechIMO.

So... one observation, and two questions...

Observation: I have an Intel i7 920, the new series of "8-core" CPUs. It is a quad-core, plus, through hyperthreading, presents the OS with 8 cores. So I am thinking, one CPU, already split into 4 execution cores, and then each core is H/T'ed into 2? I am never gonna get all-out CPU performance because so much of the CPU's power is "virtual"...

MAN was I wrong. This thing folds like there is no tomorrow. If the HT and multiple cores REALLY work like true multiple CPUs, I should get a theoretical 21 GHz out of the whole thing. So, I load it up with multiple instances of FAH (I am currently using the no-nonsense 6.23 as a service)... I mean load it ALL the way up, and guess what... It churns out Gromacs WUs at a little MORE than 10x my 2 GHz P4 single core. Actually, a fair margin more than 12x. And the heat sink always feels room temperature even though the fan never speeds up. I LOVE this processor.

Question 1. Why should I go through the trouble of SMP? Right now I have 6 individual instances of vanilla FAH, each affinity'ed to one core, churning away, and I do other stuff with the other 2 cores. What would I gain by switching over to SMP?

Question 2. I am hesitant to run the GPU FAH on this platform's ATI 4850. It would sail but this is my everyday desktop, and there is only so hard I want to push it... This is an over-the-counter box, and it may be a little lightweight... It has a 500w PSU (too light) and does not have any aftermarket cooling, so I don't want to go crazy.

But I do have my old folding platfroms still running, now 2 Dell Optiplex GX280s with single core P4s. I just set them for -advmethods and they now churn out a Gromacs P248x WU in a few days or so, which feels anemic now. So... what retail GPU has the best bang per buck? The givewaways here at TechIMO got me thinking that I should just go buy a couple of inexpensive GPUs for those two GX280s and fold on the GPU. Recommendations? I don't want to drop 250 USD into each, I'd rather find GPUs that will deliver reasonable PPD and maybe even look for them on e-bay, and just run them into the ground.

Thanks!

[KarmaKiller]

Whooa!! MOAR MEMBERS!!!



Nice to always have new people joining up.. And now on to the answers..

Question 1. Why should I go through the trouble of SMP? Right now I have 6 individual instances of vanilla FAH, each affinity'ed to one core, churning away, and I do other stuff with the other 2 cores. What would I gain by switching over to SMP?

Here's the deal on that. Most of the single core WU's are worth ~300-400points, right. Now you do have 6instances going, that's a little crazy I gotta admit, but do the math and that puts you at what, ~2100points or so, give or take some..
A single SMP WU is worth anywhere from 1760points to 1920points, depending on how you set it up. So really, with all those instances, you can probably get the same amount of PPD< with much less hassle, just using SMP.
Bizkid on our team has an I7 setup, and last he told me, it's putting out well over 3k PPD, just running a single SMP instance. There are ways to run dual SMP's on some quad cores, but you get into setting core affinity, or VMware to make that go easy. So to answer, basically you'll get the same, if not more PPD, running just SMP, with much less hassle.

Question 2. I am hesitant to run the GPU FAH on this platform's ATI 4850. It would sail but this is my everyday desktop, and there is only so hard I want to push it... This is an over-the-counter box, and it may be a little lightweight... It has a 500w PSU (too light) and does not have any aftermarket cooling, so I don't want to go crazy.

As for the 4850, you can crunch on that all day with a 500w PSU. Folding doesn't take up near the wattage as a GPU under load while gaming. I run dual 8800gt's on a FSP 350w PSU, and actually it still games almost daily, and I've yet to have a problem with it. So you could run the 4850 if you wanted to, and temps allowed of course.

Now as far as the older boxes, the catch with those is they need to have PCI-e slot. You can't fold on AGP cards. (well you can, but you have to mod the BIOS on the card, which is risky, and a PITA) so it's easier just to say you can't.
If the older boxes have PCI-e slots though, the best bang for buck folding card is the 8800gs/9600gso. And not the crappy new GSO's with only 48 steam processors, the first revision of those cards had 96 steam processors, which is the one's you looking for.

They will net ~5k a day, depending on WU. But the newer 511point WU's will bring any GPU down to it's knees, so on those they average a hair over 3k PPD. But they can be had generally for ~$50 with rebates/ebay/second hand, so they are what your looking for.

Of course, depending on PSU too...

Hmm.. did I miss anything?

[nunyadam]

Welcome to the team!
Karma the GX280's have pci-e. The question is are these the desk top units or towers? question 2 Will the Power Supplies be big enough?

I don't even want to think about a GX280 after the POS i pulled out of the dumpster.

[sharder8]

Welcome to the team Sky King!

So, is it MN or IA????

There's a couple of us MN men here on the team already . . . and we like to whoop it up a bit.

Shyguy's a southerner, while I was born and raised a northerner (then transplanted to Central Oregon).


Harder

[Sky King]

Your SMP cost/benefit analysis, assumes there's a fair PITA quotient involved... I really have none, the cores just run and run and run, I never have to touch anything. I am currently getting all 905-point P248x units, and I get through each one at about 330 PPD. When I do run all 8 cores on FAH, I yield about 2,600 PPD. I have a multi-core CPU graph on my sidebar so I can see at a glance I am running the cores at 100%. So far, I haven't touched a thing. Starting up 8 service instances in 8 folders only took maybe 2 minutes and it's a one-time thing--less time than setting up SMP would have taken. That's why I am torn about jumping in to SMP when life is pretty good running this way.

-Thx!

[Sky King]

Welcome to the team Sky King!

So, is it MN or IA????

There's a couple of us MN men here on the team already . . . and we like to whoop it up a bit.

Harder

Well, that all depends... I live in MN, in the near northern burbs... Or I should say my family does, LOL. I am an interim CEO that specializes in turn-around work, so I am currently on a longer-term gig on Cedar Rapids, IA. Woo Hoo. So IA during the week, MN on the weekends. My folding iron is an i7 and P4 in Cedar Rapids, and a P4 in the cities.

[Sky King]

Welcome to the team!
Karma the GX280's have pci-e. The question is are these the desk top units or towers? question 2 Will the Power Supplies be big enough?

I don't even want to think about a GX280 after the POS i pulled out of the dumpster.

They are both full-size towers, I can't get into one just this minute to check, I am assuming there's a few hundred watts. I used to have a pretty good fleet of 280s and they were rock-solid workhorses for me.

[2monsters]

Welcome to the team Skyking! Gpu2 client will net you the most points. Smp is really not that much of a hassle, judging by the way you have your clients set up now it should be a breeze for ya. All in all i think the SMP client will net more points than your current setup. I would also consider picking up a couple cheap gpu's as kk suggested. 2 8800GS's will get you anywhere from 5000-8000 ppd depending on wu.

[KarmaKiller]

Hmm.. well in theory, like I said before, you can run dual SMP instances. All though it may be tricky to set up, but you'll be netting the better part of 6k a day with just your CPU.

[Sky King]

Hmm.. well in theory, like I said before, you can run dual SMP instances. All though it may be tricky to set up, but you'll be netting the better part of 6k a day with just your CPU.

Are SMPs assigned higher "value per CPU tick" WUs than the 905-point FahCore_78 gigs that I am getting currently with big/advmethod on? I guess I thought that once I had big and advmethod on, that was as good as it gets in terms of what point values I have access to.

Let's see... assume I run one SMP instance and get 3K PPD... Plus I run up my 4850 and that nets another 4K PPD... Plus I run 2x 8800s and get another 8K PPD... I think 15K PPD exceeds my original goal just a bit. I will fold more in 2 days than the last 2 years.

[Bizkitkid2001]

Your SMP cost/benefit analysis, assumes there's a fair PITA quotient involved... I really have none, the cores just run and run and run, I never have to touch anything. I am currently getting all 905-point P248x units, and I get through each one at about 330 PPD. When I do run all 8 cores on FAH, I yield about 2,600 PPD. I have a multi-core CPU graph on my sidebar so I can see at a glance I am running the cores at 100%. So far, I haven't touched a thing. Starting up 8 service instances in 8 folders only took maybe 2 minutes and it's a one-time thing--less time than setting up SMP would have taken. That's why I am torn about jumping in to SMP when life is pretty good running this way.

-Thx!

When I was running two SMP instances on my Core i7 I was getting over 6,000 PPD. 6,000 > 2600.

[Sky King]

When I was running two SMP instances on my Core i7 I was getting over 6,000 PPD. 6,000 > 2600.

LOL, smart alec... OK, I am convinced, I'll give SMP a whirl. I'll let my existing WUs wrap up over the next few days and once I have 4 cores open, I'll try the SMP client. Based on other feedback here, I may run 1x SMP and 1x GPU rather than 2x SMP.

BTW, Can I go into the command lines for each of my server instances, insert the "-oneunit" flag, and then restart the services?

[nunyadam]

can't answer that, but when you set up SMP MPICH is easier from my experience. The most common problem i have seen with setting up SMP is people forgetting to insert a "SPACE" before the -smp command.

I need to do some research, or somebody post me a link as to how to get SMP and Vista working. its probably something silly I'm missing, but my Brother inlaws AMD machine kept giving me errors when i tried to set it up last night.

[KarmaKiller]

LOL, smart alec... OK, I am convinced, I'll give SMP a whirl. I'll let my existing WUs wrap up over the next few days and once I have 4 cores open, I'll try the SMP client. Based on other feedback here, I may run 1x SMP and 1x GPU rather than 2x SMP.

BTW, Can I go into the command lines for each of my server instances, insert the "-oneunit" flag, and then restart the services?

I *think* you can, but honestly I've never tried it before.
I myself never run any of them as service, it's easier for me to control everything that way.

[pullmyfoot]

As for the 4850, you can crunch on that all day with a 500w PSU. Folding doesn't take up near the wattage as a GPU under load while gaming. I run dual 8800gt's on a FSP 350w PSU, and actually it still games almost daily, and I've yet to have a problem with it. So you could run the 4850 if you wanted to, and temps allowed of course.

Wow you are insane. I thought you were running just one

[KarmaKiller]

Wow you are insane. I thought you were running just one

Well, that's still up for debate in the high courts.. for now, no answer.

(I'm sure the others will chime in pretty quick)

[Bizkitkid2001]

LOL, smart alec... OK, I am convinced, I'll give SMP a whirl. I'll let my existing WUs wrap up over the next few days and once I have 4 cores open, I'll try the SMP client. Based on other feedback here, I may run 1x SMP and 1x GPU rather than 2x SMP.

BTW, Can I go into the command lines for each of my server instances, insert the "-oneunit" flag, and then restart the services?

You could also run 2x SMP and the GPU. I currently am running 1x SMP for testing reasons, but if I were to run 2x SMP and GPU I would be getting over 13k PPD.

I am running 1 SMP now because of problems me and Karma ran into with Vista 64. But I think Karma finally found a fix.

[Sky King]

You could also run 2x SMP and the GPU. I currently am running 1x SMP for testing reasons, but if I were to run 2x SMP and GPU I would be getting over 13k PPD.

I'll be curious to see the fix, as I too am running Vista 64 (ultimate).

I loaded up the GPU client tonight, and that is kicking out about 3,200 PPD by itself, so my current setup is 6x no-nonsense consoles running as big/advmethods services, and 1x GPU, for a single-chassis total of 5,200 PPD.

My reason for thinking I'd only run 1x SMP was because I was told you really needed to dedicate a core to managing the GPU folder... The process associated with the GPU folder runs one core @ 100% on my machine...

So I was assuming "best practice" on 8 cores would be to run 1x GPU, 1x SMP, and then 3x no-nonsense affinity'ed to each of the remaining 3 cores. If 2x SMP still works with GPU, I'll have a go once I see what the Vista 64 problem and resolution are.

[KarmaKiller]

Really, the only way I know to run dual SMP's on any OS is with VMware. If it's a dedicated folder box, then really it's a better way to go. But if you use your PC for other then folding, it can be a PITA to run it like that.
But one Vmware SMP will only use 2 cores. So on Quads (without HT) you can run dual instances. I honestly don't know if VMware picks up on HT or not, but I would imagine it would, so you could probably run 4 instances.

Any of you guys wanna send me an I7 so I can test it for ya?

[Sky King]

And then the troubles begin...

I installed the GPU client, ran it, and it hums along... works through 411 point cores in about 3 hours. It finishes up, returns the work, and downlaods new work... but the process of instantiating the new execution core fails with an onerous sounding error. (see last 3 lines of log snippet:

[18:24:18] + Processing work unit
[18:24:18] Core required: FahCore_11.exe
[18:24:18] Core found.
[18:24:18] Working on queue slot 03 [March 10 18:24:18 UTC]
[18:24:18] + Working ...
[18:24:18] - Calling '.\FahCore_11.exe -dir work/ -suffix 03 -checkpoint 12 -verbose -lifeline 5880 -version 623'
[18:24:18] 
[18:24:18] *------------------------------*
[18:24:18] Folding@Home GPU Core - Beta
[18:24:18] Version 1.22 (Mon Dec 8 12:57:56 PST 2008)
[18:24:18] 
[18:24:18] Compiler  : Microsoft (R) 32-bit C/C++ Optimizing Compiler Version 14.00.50727.762 for 80x86 
[18:24:18] Build host: amoeba
[18:24:18] Board Type: AMD
[18:24:18] Core      : 
[18:24:18] Preparing to commence simulation
[18:24:18] - Looking at optimizations...
[18:24:18] - Created dyn
[18:24:18] - Files status OK
[18:24:18] - Expanded 88215 -> 447304 (decompressed 507.0 percent)
[18:24:18] Called DecompressByteArray: compressed_data_size=88215 data_size=447304, decompressed_data_size=447304 diff=0
[18:24:18] - Digital signature verified
[18:24:18] 
[18:24:18] Project: 4744 (Run 6, Clone 10, Gen 12)
[18:24:18] 
[18:24:18] Assembly optimizations on if available.
[18:24:18] Entering M.D.
[18:24:28] CoreStatus = FFFFFFF6 (-10)
[18:24:28] Client-core communications error: ERROR 0xfffffff6
[18:24:28] This is a sign of more serious problems, shutting down.

When I manually restart the systray client, it plowed on and completed the work unit. Could I be having problems with Vista UAC impeding something? That seems odd, as any process talking to its own spawned subprocess is non-privileged communication.

Recommendations?