folding on skulltrail

[Insidious]

Ah GLeeM

You are such an optimist!



But it is a good idea to use those guys (actually, some of them are great!).

However, I would mention that I have NEVER seen a problem like this introduced over there and get any answer but "your hardware needs tested". (They get their feelings hurt easily if they are questioned about the ... perfectness of their software.)

So I'd wait until you have some hardware test results to post in your original request. Otherwise they probably aren't going to be very helpful ('magin that)

-Sid

 

[biodoc]

Derek,

We really do appreciate an Anandtech Editor mixing it up with us DC folk. We really do!

I think Sid was the first to suspect it's a hardware issue with your test machine.

A couple of points:

1) the fah504 console client is incredibly stable. There are times when I'm so frustrated with the SMP client on Vista, I just delete the directory and run 4xfah504 on my quad for a break.

2) Most of us are drooling for an 8 core system so we "doubted" the stability of the 504 client on such a marvelous piece of hardware.

3) I'm an impatient guy when it comes to long hardware tests but they are probably useful at this time.

4) However, I would just pull half those stick of RAM out, test fah, then substitute one stick for the other to test the RAM my way.

I suspect Sid was right in his first post, it's a bad stick of RAM.

If not, then it's a more serious hardware problem.

Anyway, we're glad you're here with us DC enthusiasts and one thing for sure....if you truly want to test your hardware, run fah!

 

[DerekWilson]

thanks again guys ... i'll have more time tomorrow to mess with this stuff ...

i have been running latest bios and drivers and all that, except i haven't installed vista sp1 yet. just running hotfixes.

it could be a hardware issue. but like i said, i won't have my new fbdimms for a while, so if its one of those i'm stuck for a while and that's bad ... especially because i need to test a new ... uhh ... thing ... from ... a graphics card maker ... on the skulltrail box next week.

 

[DerekWilson]

just went and checked a couple things ...

5 clients seems to run fine still ... but i think i found something ... I powered it down and was gonna pull one memory module ... but it burnt my fingers ...

memory ain't aposedta get dat hot ...

part of the issue could be that i've not got this thing in a case. and there aren't any fans blowing across anything ... though the gaint ass CPU fans looked like they might help, i suppose they are not.

so now i've got a 120 mm fan blowing down on the ram from 5 inches away... i'm just gonna try running all 8 clients again ... if it fails i'll hit the memtest and gromacs test stuff ... but i think this has a shot at working.

well see ... gnight for now.

.... ... by the way, sorry i can't spend a lot more time working on this at the moment ... i'm going as fast as I can

 

[Insidious]

Hobbies wouldn't be any fun if they didn't take a while.

I think it's cool that you're hangin' in with this! :thumbsup:

-Sid

hmmmm.......! new graphics gizmo that needs a skulltrail?

sounds FUN!

If you have any difficulty setting up 4 GPU clients... just let us know. We probably wouldn't be much help, but would be happy to meddle.

 

[DerekWilson]

guess what's still going strong after 2 hours ... this is about the breaking point i've seen in the past ... everyone hold your breath

if it fails i've got memtest86 ready to run ... we'll find out if its memory tonight at least. if memtest doesnt fail out i'll switch to the gromacs test ...

 

[caferace]

Originally posted by: Insidious
We probably wouldn't be much help, but would be happy to meddle.

That, right there, should be the TeAm DC Motto. :thumbsup: :laugh:

-jim

 

[DerekWilson]

it's kind of a shame that i only have 2 1950xtx cards laying around ... i'd love to plug three of em in there (plus one single slot x1900orother...)

they need to get it working on the r600 and g80

but that's a whole other thread

 

[DerekWilson]

.... still going ... ... ...

it was a ram cooling issue ...

i've actually never run into that on a stock clocked system with loose memory timings ... that's pretty crazy to me.

 

[DerekWilson]

... now to try running 2 GPU clients, and 3 SMP clients

... actually i've got to spend time with my family first ... been doin too much work and not enough of that lately ... this stuff is fun but it still takes second to my wife and daughter

 

[Insidious]

Glad to hear that it was nothing more than that on your skulltrail.

Don't forget that each GPU client will want to have a CPU to work with. You may run into some resource limitations trying to run 12 SMP threads at the same time as 2 GPU clients. I recommend using affinity settings to keep those SMPs limited to 6 of your cores and the GPUs set, one each to the remaining two.

-Sid

edit: You may have to re-set affinity each time a client begins a new work unit

 

[Markfw]

I would like to see someone try this on XP pro with a skulltrail. I have 6 quads running 2 x smp units, and 8 threads. All the boxes have 2 gig ram. They are all overclocked to their max (under F@H, which is less than a normal OC) And they run for MONTHS with no problems 24/7 and no re-boots. Skulltrail with 8 single threads shouldn't be a problem.

The other idea would be to try 4 x SMP clients. I gave no switches set, just two things other than default, allow large units, and I set the machineid on each client. Also, I fire up in my startup, 2 different shortcuts, that have 2 different windows. I saw where he fired up
cd 1
fire 1
cd 2
fire 2
.....
That may be a problem ????? Use different console windows ???

Edit: I see you found the memory thing. I have 120 mm fans blowing on NB and memory in every box (except the Antec 900 cases do fine)

Running F@H 24/7@100% sure can stress a system, huh Derek ! And nobody else around here believes us, when we say that is the ultimate stability tester.....

And PPD wise, those GPU client will kill your overall PPD. Use FAHMON to tell you.. But then again, you aren't running the SMP client. You should be getting 9000 PPD out of a Skulltrail minimum.

 

[Foxery]

Good to hear that it was an easy fix!

Originally posted by: DerekWilson
... now to try running 2 GPU clients, and 3 SMP clients

Remember that SMP clients may miss deadlines when you need to stop them for your day job But the results will certainly be interesting.

For what it's worth... there's a tool called AffinityChanger which helps people on quad-core chips, BUT I'm not sure if it will detect/properly deal with eight.

As for the GPU client, I'd actually suggest waiting for the new client and experimenting with 3000-series cards instead, if possible. Given the reason why you have this rig, that probably isn't a stretch of the imagination

 

[DerekWilson]

actually, when r600 (or g80/g90) get support i'm planning on doing an article including performance relative to r500, x86, ppc, and PS3 ... as i understand it, the gpu core and the ps3 core target specialized work units and can't handle all the stuff a normal CPU handles ... i'll have to figure out how to compare their processing ability using as similar a work unit as possible ... which means i'll likely have to talk to the people at stanford about it ...

but they seem to take a while to add support for things ..

 

[GLeeM]

Originally posted by: DerekWilson
as i understand it, the gpu core and the ps3 core target specialized work units and can't handle all the stuff a normal CPU handles..

Something about implicit and explicit. One would run like on a CPU the other runs 30-40 times faster.

Originally posted by: DerekWilson
... i'll have to figure out how to compare their processing ability using as similar a work unit as possible ... which means i'll likely have to talk to the people at stanford about it ...

There are many different types of WU just for CPU. Points per day on a P4 can range from 100 to over 450.

 

[Insidious]

That would be like trying to compare how well does a bowler do versus a golfer when each are playing 'similar' versions of their games.....

There is no golf course that is 'simiilar' to a bowling alley

-Sid

 

[biodoc]

Glad to hear the 8 core was an easy fix!

My understanding is the PS3, GPU & CPU are all performing quite different scientific tasks so it would really be difficult to compare them. Perhaps someday though.

 

[Rattledagger]

Originally posted by: Markfw900
Running F@H 24/7@100% sure can stress a system, huh Derek ! And nobody else around here believes us, when we say that is the ultimate stability tester.....

Hmm, finished a CPDN-wu lately... ?

 

[Markfw]

Originally posted by: Rattledagger

Originally posted by: Markfw900
Running F@H 24/7@100% sure can stress a system, huh Derek ! And nobody else around here believes us, when we say that is the ultimate stability tester.....

Hmm, finished a CPDN-wu lately... ?

Whats that ?

 

[Insidious]

CDPN is a climate prediction DC project.

The work units are very long and only rarely complete. One false bit........ pow!

-Sid

 

[petrusbroder]

Originally posted by: Markfw900

Originally posted by: Rattledagger

Originally posted by: Markfw900
Running F@H 24/7@100% sure can stress a system, huh Derek ! And nobody else around here believes us, when we say that is the ultimate stability tester.....

Hmm, finished a CPDN-wu lately... ?

Whats that ?

Climate PreDiction Net:
A BOINC-project that calculates different models for the development of the climate on earth. Takes into consideration quite many temperatures (air, oceans, ground etc.) in different locations, air movements (winds), solar input, etc., etc., etc.
A WU usually takes somewhere between 300 - 3000 hours to calculate.
The WUs are very sensitive to errors, so often OC'ing, high temperatures, variation in voltages, and any other instability results in an computation errors and an invalid result.
Of course you get credits for the time you crunch, as every other day or so a part of the result is sent to the project server, and also because your computer sends "trickle" messages to the server.
I have a success rate of approx. 60%. Considering the age of my computers etc I am happy with that result.
I find CPDN a very good - although somewhat too sensitive for normal use - stability tester. If a computer has finished a result, then it is good.
IMHO, if you really want to see if your computer calculates correctly, the CPDN is the best test.

 

[Insidious]

if you really want to see if your computer calculates correctly, the CPDN is the best test

I don't necessarily agree that the CDPN software is so infalable as to blame every failure of a work unit on the PC.
My experiences with that project led me to believe you need a VERY stable PC.... AND also you need some good luck on work unit assignment to actually complete one.

I would go so far to agree that it is a very good first indicator of a potential instability, but I think that needs to be verified by a second source before you start replacing hardware or changing too many settings.

but this is just my opinion and I have been wrong lots of times before

-Sid

 

[Rudy Toody]

Originally posted by: Insidious
but this is just my opinion and I have been wrong lots of times before

I have never been wrong before. This is the first time.

 

[Amaroque]

Originally posted by: DerekWilson
.... still going ... ... ...

it was a ram cooling issue ...

i've actually never run into that on a stock clocked system with loose memory timings ... that's pretty crazy to me.

Welcome to real stress testing! I crunch a different project for AT, but been following the thread.

Thing is, the benchmarks and that you guys do are great for the average gamer, but not for a DCer. Two completely different animals.

Glad you found the problem. :thumbsup:

 

[mrwizer]

Originally posted by: Insidious

if you really want to see if your computer calculates correctly, the CPDN is the best test

I don't necessarily agree that the CDPN software is so infalable as to blame every failure of a work unit on the PC.
My experiences with that project led me to believe you need a VERY stable PC.... AND also you need some good luck on work unit assignment to actually complete one.

I would go so far to agree that it is a very good first indicator of a potential instability, but I think that needs to be verified by a second source before you start replacing hardware or changing too many settings.

but this is just my opinion and I have been wrong lots of times before

-Sid

Your correct, CPDN units can fail from the starting variables, not just system stability.

Edit: I am doubting myself that they fail on variables too. I remember seeing that somewhere, but after I posted it I wanted to go check. Yet cant find anything definitive. Maybe Rattle can weigh in. So take my comment with a grain of salt.