Benchmark settings? New A4/SMP Projects p7645, 7646, 7647

[ob1]

Hello to everybody.

Following an email discussion with TJ, he asked me to post our messages on the forum.
"Dont acte"

Hello.

(...)

Then, I'd like to talk to you about a problem I have with the Project stated in the subject of this mail : 764x. I could have written on the forum, but I'm not in the beta team and don't feel implicated enough to be part of it.
But, with the v7 client, the GPU downloads some neats protein and folds them quite efficiently (~5 min/frame, ATI HD 5750). But with the SMP (Core 2 Quad Q6600), i've downloaded twice a p764x project and it is impossible to meet the expiration line : just now, I've been folding for 30 minutes and haven't achieved any frame !!!
Yet the FahCore_a4 process runs at a steady 75% (the remaining 25% are for the CPU part of the GPU folding).
Don't get me wrong : I don't care about PPD. I just care about folding and making our comprehension of this process bigger. But I don't see the point in using my CPU for something that 1) will be long and 2) will never ever serve you.

So, here it is : what can I do to fold these projects (7645, 7646) faster,
or fold other projects ?
Here's the log of the smp :

*********************** Log Started 2012-07-12T16:12:37Z ***********************
16:12:37:WU01:FS01:Starting
16:12:37:WU01:FS01:Running FahCore: "C:\Program Files (x86)\FAHClient/FAHCoreWrapper.exe"
C:/Users/Olivier/AppData/Roaming/FAHClient/.../AMD64/Core_a4.fah/FahCore_a4.exe -dir 01 -suffix 01 -version 701 -lifeline 3960 -checkpoint 15 -np 4
16:12:37:WU01:FS01:Started FahCore on PID 2444
16:12:37:WU01:FS01:Core PID:3036
16:12:37:WU01:FS01:FahCore 0xa4 started
16:12:38:WU01:FS01:0xa4:
16:12:38:WU01:FS01:0xa4:*------------------------------*
16:12:38:WU01:FS01:0xa4:Folding@Home Gromacs GB Core
16:12:38:WU01:FS01:0xa4:Version 2.27 (Dec. 15, 2010)
16:12:38:WU01:FS01:0xa4:
16:12:38:WU01:FS01:0xa4:Preparing to commence simulation
16:12:38:WU01:FS01:0xa4:- Looking at optimizations...
16:12:38:WU01:FS01:0xa4:- Files status OK
16:12:38:WU01:FS01:0xa4:- Expanded 548213 -> 848972 (decompressed 154.8 percent)
16:12:38:WU01:FS01:0xa4:Called DecompressByteArray: compressed_data_size=548213 data_size=848972, decompressed_data_size=848972 diff=0
16:12:38:WU01:FS01:0xa4:- Digital signature verified
16:12:38:WU01:FS01:0xa4:
16:12:38:WU01:FS01:0xa4:Project: 7645 (Run 488, Clone 0, Gen 10)
16:12:38:WU01:FS01:0xa4:
16:12:38:WU01:FS01:0xa4:Assembly optimizations on if available.
16:12:38:WU01:FS01:0xa4:Entering M.D.
16:12:44:WU01:FS01:0xa4:Mapping NT from 4 to 4
16:13:27:WU01:FS01:0xa4:Completed 0 out of 2500000 steps (0%)

PS : I've tried client-type=advmethods and got these molecules again.

He kindly answered this :

Olivier,

Thanks for this report, and your continued dedication to folding! I completely see your point here, that we are basically wasting your valuable donated CPU, which of course is not good. There are a few options here that I will look into, which include

(1) increasing the deadline, so that donors such as yourself have more time to complete the WUs
(2) decreasing the length of each WU
(3) restricting these WUs to donors running with more powerful computers

Let me do some digging around and decide which is best for the project and the donors. If you have input, sent it along! Then we'll fix this in the next day or so - sorry for all the lost time on your machine, definitely not what either of us want!

Best,

TJ

PS. Feel free to post such messages to the forum next time. That way, other people with the same problem might also report it up. Without feedback from you guys, we have very little idea what issues arise, so these reports are very helpful.

Then I saw the this project would be reserved for 4+ cores.

Has any of you got this problem ?

PS : I've changed my smp configuration, enabling just 2 cores instead of 3 (well, -1, so the whole 4 cores minus the one used for the GPU). I got the 7611 project (3, 56, 131), and it is, again, insanely slow

Luckily, my GPU folds solidly at 4'24 / frame

[7im]

Hello ob1, welcome to the forum.

FYI, the V7 client for Windows is no longer considered a beta client by the Pande Group. You don't need to be a beta tester to discuss that client here in the forum.

In regards to 4+ cores, the assignment server only looks at the total number of cores you have, not how many are configured for use in the client (3, or now 2). A lot of people will have that problem. Hopefully TJ read this and chooses option 3.

Maybe that needs to become a feature request. The Assignment Server (AS) should use the lesser of the two values, the System core count and the number of threads configured in the client.

[ob1]

Thank you very much indeed. Just restarted and got 8041 (16, 161, 21), folding at ~9'/f

[bruce]

Also please note that the beta testing of the client has been kept separate from the beta testing of new projects and new FahCores that might be required to support them.

> When any new FahCore is being tested, it is not distributed to the general public so others will not encounter any bugs that are discovered by the BetaTeam. We ask that all discussions be kept inside the BetaTeam Forum and we ask that only those with access to that forum configure their hardware to receive WUs using this revision.
> When any new project is initially made available, it is distributed first to beta team members who are experienced in reporting problems with particular WU or with the FahCore it uses. We ask that all discussions be kept inside the BetaTeam Forum and we ask that only those with access to that forum configure their hardware to receive these WUs.
> Once a new project has completed the early phase of beta testing, it migrates to Advanced Methods. Anyone can select this option and discussions are unrestricted. The projects may still experience a higher failure rate that projects that have been released to full FAH status.
> New clients are introduced much less often. Once a beta version is ready to be distributed information about it is posted in a public forum and anyone can participate in that testing. Several V7 beta versions have been tested publicly and the Windows and Linux versions are now recommended on the folding homepage. The OS-X beta version isn't as mature yet but it can still be found if you want to beta test it. Development of the V7 client is still on-going, but at a significantly slower pace.

Beta testing doesn't always find every problem so reports like yours are always welcome at any stage of development.

[Joe_H]

I have been folding these projects since they went to advanced out of beta. They definitely are large WU's, so if you are used to small ones that take a day or less I can see them not being suitable. But, even my MacBook with a 2.4 Core 2 Duo can complete them well within the preferred time. In fact it is working on a 7647 that it downloaded earlier this week that will turn in on Saturday. That takes about 4.5 days of the 10.4, so I have to assume your folding includes a number of periods when you have paused the client and gone over the deadline.

I have also folded a number of these on my iMac and my son's Mac Pro. The iMac takes about 33-34 hours to complete one, the Mac Pro slightly less when folding at SMP8. Currently my son has it at SMP6 to leave some processing available for gaming, the time it takes to complete is around 40 hours.

As for improvements in the assignment algorithm so people get work assigned that they feel is more appropriate for their computer, that has been discussed many times. Hopefully some of the proposals eventually get implemented that are consistent with the needs of the projects. I don't have any favorites myself, but can see a need for some changes eventually.

P.S. Just saw the notice from tjlane of the move from advanced to FAH for these projects. The change to a minimum of SMP4 will be useful as I can fold them on my MacBook, but prefer to fold smaller WU's on it for ease of scheduling when I stop folding for extended periods when I will be running off the battery.

[Jesse_V]

Maybe that needs to become a feature request. The Assignment Server (AS) should use the lesser of the two values, the System core count and the number of threads configured in the client.

Makes sense to me.

[bruce]

I've changed my smp configuration, enabling just 2 cores instead of 3 (well, -1, so the whole 4 cores minus the one used for the GPU). I got the 7611 project (3, 56, 131), and it is, again, insanely slow

Luckily, my GPU folds solidly at 4'24 / frame

I'm not sure I completely understand. Your log shows SMP only and a configuration of SMP:4 (which is what you'll get if you specify SMP:-1). It doesn't show that you're folding with a GPU.

Assuming you've somehow separated the logs and there's another log somewhere for a GPU that's folding concurrently, I suggest you manually reconfigure your SMP slot for SMP:3. That will be more productive that either SMP:2 or SMP:4. You should see Mapping NT from 3 to 3 rather than Mapping NT from 4 to 4

[ob1]

Your log shows SMP only and a configuration of SMP:4 (which is what you'll get if you specify SMP:-1). It doesn't show that you're folding with a GPU.

I've selected only the smp slot logs. The GPU is actually folding :

Code: Select all

... 04:51:22:WU00:FS00:Upload 81.58%
04:51:28:WU00:FS00:Upload complete
04:51:28:WU00:FS00:Server responded WORK_ACK (400)
04:51:28:WU00:FS00:Final credit estimate, 1835.00 points
04:51:28:WU00:FS00:Cleaning up
04:55:26:WU02:FS00:0x16:Completed    500000 out of 50000000 steps (1%).
05:00:14:WU02:FS00:0x16:Completed   1000000 out of 50000000 steps (2%).
05:05:02:WU02:FS00:0x16:Completed   1500000 out of 50000000 steps (3%).
05:09:54:WU02:FS00:0x16:Completed   2000000 out of 50000000 steps (4%). ...

I suggest you manually reconfigure your SMP slot for SMP:3.

When I configure the smp slot, there's an hint that says :

(The number of CPU thread this slot uses) should be a multiple of 2.

So I put either 2 or 4. But maybe I don't understand it very well ...

[tjlane]

Hey All,

Earlier today I restricted these projects to clients running with 4+ cores (http://foldingforum.org/viewtopic.php?f=24&t=21654). Hopefully this alleviates the problem. Please don't hesitate to report here with continued deadline issues, and we can continue to adjust accordingly.

Thanks,

TJ

[Zagen30]

I suggest you manually reconfigure your SMP slot for SMP:3.

When I configure the smp slot, there's an hint that says :

(The number of CPU thread this slot uses) should be a multiple of 2.

So I put either 2 or 4. But maybe I don't understand it very well ...

Some projects have major problems when running on a prime number of cores due to issues in breaking up a WU into that many pieces for processing. The exceptions are 2, 3, and most likely 5 (I don't believe it's ever been officially confirmed that 5 is trouble-free, but I can't recall anyone reporting problems that were due to running smp:5). 7 is the most common trouble spot since a lot of people have i7's and have tried to run on 7 of the 8 threads they have to leave one thread available for the GPU. Suggesting an even number means that you won't unknowingly enter a problem number, but that's why they used the word "should" instead of "must."

[csvanefalk]

I am not here for the PPD either, but is not p7645 terribly undervalued? On my rig, I get a TPF of about 17 minutes (!!!), and my PPD drops to about 15k. Normally, I see PPDs ranging from 40-70k

[Joe_H]

On my i7-860 iMac running these I get about 11.4k PPD before the current 10% bonus for A4's is added. The TPF is about 20 minutes. That is about 10% less than I see from some of the projects in the 8000's, but not greatly less. What projects are you seeing as much as 40-70k for PPD? Your 3930K has 6 cores to my 4 and a stock clock of 3.2 GHz compared to 2.8 GHz, so a TPF of 17 minutes sounds high. I would check on what else is running and taking CPU time.

[iceman1992]

Lower yes, but I wouldn't say terribly undervalued.. I have a 7646 running with a TPF of 17:45, est PPD about 13500 without 10% bonus, and I'm using a 4C/4T i5.

[artoar_11]

I am not here for the PPD either, but is not p7645 terribly undervalued? On my rig, I get a TPF of about 17 minutes (!!!), and my PPD drops to about 15k. Normally, I see PPDs ranging from 40-70k

The same projects with i5 2500K @ 4.20 GHz; VM/Ubuntu 12.04/64b:

p7645 - TPF 15:18; PPD ~17 100
p7646 - TPF 15:27; PPD ~16 800
p7647 - TPF 15:18; PPD ~17 100

This is according "Bonus Point Calculator". Must be add +10%. Normally I receive - 20-22k PPD on this frequency (with core A3).

[iceman1992]

artoar_11, how do you get TPF of 15 minutes?? I have a very similar setup, i5-2500K @ 4.1 GHz, VM ubuntu 12.04 64bit, and I get a TPF of more than 17.5 mins.. The difference of 0.1 GHz surely can't give more than 2 minutes??