Folding Forum

Posted: **Wed Dec 26, 2007 2:47 pm**

Hi,
I've recently started getting Project 3062 for my linux SMP client and have noticed my PPD has decreased on average about 600 points. I understand that different hardware will net different frame times but I am running a Q6600 at standard clock frequency. I believe that this is the same reference processor used by Stanford to calculate the points given to a work unit so shouldn't my PPD not differ greatly no matter what work unit I receive? All other units I've completed have hovered around the 2400 PPD mark.
I've received this project about four times now and I'm still getting a hit in my PPD. I've restarted my computer as well to see if anything was interfering but to no avail. The project also isn;t on fahinfo.org so I've no way of checking what others frame times are, and also no one on my team has encountered this project on similar hardware. Am I missing something obvious here? (I do that a lot, so educate me please!) Thanks for taking the time to read

Here's my most recent log file of this project. Frame times according to FahMon are averaging: 10min 08sec

Code: Select all

--- Opening Log file [December 26 12:28:42] 


# SMP Client ##################################################################
###############################################################################

                       Folding@Home Client Version 6.00beta1

                          http://folding.stanford.edu

###############################################################################
###############################################################################

Launch directory: /home/david/foldingathome/CPU1
Executable: /home/david/foldingathome/CPU1/fah6
Arguments: -smp -verbosity 9 

[12:28:42] - Ask before connecting: No
[12:28:42] - User name: Oak37 (Team 52286)
[12:28:42] - User ID: 66B7ACE8192C949D
[12:28:42] - Machine ID: 1
[12:28:42] 
[12:28:42] Loaded queue successfully.
[12:28:42] 
[12:28:42] + Processing work unit
[12:28:42] Core required: FahCore_a1.exe
[12:28:42] Core found.
[12:28:42] - Autosending finished units...
[12:28:42] Trying to send all finished work units
[12:28:42] + No unsent completed units remaining.
[12:28:42] - Autosend completed
[12:28:43] Working on Unit 04 [December 26 12:28:43]
[12:28:43] + Working ...
[12:28:43] - Calling './mpiexec -np 4 -host 127.0.0.1 ./FahCore_a1.exe -dir work/ -suffix 04 -checkpoint 30 -verbose -lifeline 5720 -version 600'

[12:28:43] 
[12:28:43] *------------------------------*
[12:28:43] Folding@Home Gromacs SMP Core
[12:28:43] Version 1.74 (November 27, 2006)
[12:28:43] 
[12:28:43] Preparing to commence simulation
[12:28:43] - Ensuring status. Please wait.
[12:29:01] - Looking at optimizations...
[12:29:01] - Working with standard loops on this execution.
[12:29:01] - Previous termination of core was improper.
[12:29:01] - Going to use standard loops.
[12:29:01] - Files status OK
[12:29:01] - Expanded 608147 -> 3255645 (decompressed 535.3 percent)
[12:29:01] 
[12:29:01] Project: 3062 (Run 4, Clone 17, Gen 5)
[12:29:01] 
[12:29:01] Entering M.D.
[12:29:07] Protein: p3062_lambda5_99sb
[12:29:07] Writing local files
[12:29:07] Completed 3450000 out of 5000000 steps  (69 percent)
[12:29:07] Extra SSE boost OK.
[12:29:07] 
[12:29:07] Completed 3450000 out of 5000000 steps  (69 percent)
[12:29:07] Extra SSE boost OK.
[12:40:06] Writing local files
[12:40:06] Completed 3500000 out of 5000000 steps  (70 percent)
[12:51:11] Writing local files
[12:51:11] Completed 3550000 out of 5000000 steps  (71 percent)
[13:02:10] Writing local files
[13:02:10] Completed 3600000 out of 5000000 steps  (72 percent)
[13:13:05] Writing local files
[13:13:05] Completed 3650000 out of 5000000 steps  (73 percent)
[13:24:01] Writing local files
[13:24:01] Completed 3700000 out of 5000000 steps  (74 percent)
[13:35:03] Writing local files
[13:35:03] Completed 3750000 out of 5000000 steps  (75 percent)
[13:46:06] Writing local files
[13:46:06] Completed 3800000 out of 5000000 steps  (76 percent)
[13:57:05] Writing local files
[13:57:05] Completed 3850000 out of 5000000 steps  (77 percent)
[14:08:06] Writing local files
[14:08:06] Completed 3900000 out of 5000000 steps  (78 percent)
[14:19:05] Writing local files
[14:19:05] Completed 3950000 out of 5000000 steps  (79 percent)
[14:30:10] Writing local files
[14:30:10] Completed 4000000 out of 5000000 steps  (80 percent)
[14:41:11] Writing local files
[14:41:11] Completed 4050000 out of 5000000 steps  (81 percent)

note: The previous improper termination was me resetting the computer after it froze turning off. Forceasm is on by default so standard loops shouldn't be playing a part in this.

Posted: **Wed Dec 26, 2007 5:38 pm**

p3061 and p3062 are quite a bit slower than all other WUs on the Q6600 for some reason. For reference, my comparative results from a Q6600 @ 3.33 GHz, 370 x 9, 4:5, 4-4-4-12, Ubuntu 7.10.

Code: Select all

	Project				              Pts.	     TPF		  TPU		 PPD
2653 (Run 2, Clone 25, Gen 24)	  1760	  06:53.2	  11:28:38	  3680.3
3043 (Run 9, Clone 17, Gen 61)	  1440	  06:15.1	  10:25:06	  3317.3
3050 (Run 1, Clone 6, Gen 79)		1440	  06:28.8	  10:48:00	  3200.0
3051 (Run 5, Clone 2, Gen 75)		1440	  06:24.1	  10:40:10	  3239.2
3051 (Run 5, Clone 2, Gen 75)		1440	  06:27.3	  10:45:30	  3212.4
3052 (Run 0, Clone 8, Gen 70)		1440	  06:21.6	  10:35:57	  3260.6
3059 (Run 4, Clone 89, Gen 42)	  1167	  05:08.2	   8:33:40	  3271.5
3060 (Run 11, Clone 12, Gen 2)	  2539	  10:57.9	  18:16:30	  3334.4
3061 (Run 0, Clone 22, Gen 4)		1329	  07:20.7	  12:14:30	  2605.5
3062 (Run 5, Clone 6, Gen 10)		1324	  07:20.7	  12:14:30	  2595.7

p3061 and p3062 produce about 27% less than the average of the other WUs I've recorded results for on this machine.

Posted: **Wed Dec 26, 2007 10:22 pm**

Thanks for replying, it's good to know it's not just me! Maybe the points for the unit should be increased if this is a prevalent issue?

Posted: **Thu Dec 27, 2007 8:42 am**

Oak37 wrote:Thanks for replying, it's good to know it's not just me! Maybe the points for the unit should be increased if this is a prevalent issue?

The SMP benchmark is about 190 PPD/GHz/core. It's not as helpful to compare one project to another, some slightly underperform, some over, some on the mark, depending on your specific hardware. Let's start with what PPD you are actually getting, and let's go from there...

Posted: **Thu Dec 27, 2007 5:34 pm**

7im wrote:The SMP benchmark is about 190 PPD/GHz/core. It's not as helpful to compare one project to another, some slightly underperform, some over, some on the mark, depending on your specific hardware. Let's start with what PPD you are actually getting, and let's go from there...

The PPD average running p3062 is 1875 points. I know that some units will under perform while others will do the opposite but what grabbed my attention was the deviation from my average of 2400 PPD. I thought this was strange considering that I'm running the same processor as the reference processor used.

Posted: **Thu Dec 27, 2007 9:59 pm**

Is the benchmark machine in fact a Q6600? If it is the disparity would be hard to explain. If the benchmark machine were say a dual Woodcrest @ 2.4, the Q6600 may outperform the dual Woodcrest by 27% on most WUs, but have little to no advantage on some like p3061 and p3062 ( and p2652). What are the specs of the SMP benchmark machine?

Posted: **Thu Dec 27, 2007 10:42 pm**

The point system can be confusing sometimes. Some projects get bonus points, so that automatically increases their PPD.
Does 3062 use less memory than, say, 2653?

Dr Pande has promised to revamp the point system soon, so there might be solution in sight.

Posted: **Thu Dec 27, 2007 11:42 pm**

ChasR wrote:Is the benchmark machine in fact a Q6600? If it is the disparity would be hard to explain. If the benchmark machine were say a dual Woodcrest @ 2.4, the Q6600 may outperform the dual Woodcrest by 27% on most WUs, but have little to no advantage on some like p3061 and p3062 ( and p2652). What are the specs of the SMP benchmark machine?

RTFree SMP FAQ. How do you decide the credit value of SMP work units? i.e. dual Woodcrests @ 2.33 GHz.

Posted: **Fri Dec 28, 2007 11:27 am**

7im wrote:RTFree SMP FAQ. How do you decide the credit value of SMP work units? i.e. dual Woodcrests @ 2.33 GHz.

Thanks for that, I remember reading somewhere a long time ago that the reference processor was a Q6600, but now I know it's not, so it explains the points disparity

Posted: **Fri Dec 28, 2007 2:16 pm**

Oak37 wrote:
7im wrote:RTFree SMP FAQ. How do you decide the credit value of SMP work units? i.e. dual Woodcrests @ 2.33 GHz.
Thanks for that, I remember reading somewhere a long time ago that the reference processor was a Q6600, but now I know it's not, so it explains the points disparity

I'm not sure that I understand the reasoning here.

Depending on the motherboard/chipset being used with the Q6600,
they should be about equal, with the Q6600 coming out ahead by a small bit.

The Q6600 is, after all, two Woodcrests, on one package, yes?

As long as the work stays within the L2 cache, it should be completed in about the same time.

This does not, however, explain the discrepancy in points, when the Q6600 is run @ 3.2GHz, or more,
and still performs at a level of 600 - 700 ppd less.

Refer to ChasR's very accurate chart, above for examples.
This is very repeatable and agrees with the reports from my teammates, and others.

Posted: **Fri Dec 28, 2007 3:11 pm**

Either the WUs aren't benched correctly, have been given a smaller point value for a reason, or most likely, the Q6600 on the p965, p975 or p35 chipsets outperforms the dual Woodcrests by a 25%+ margin on most, but not all WUs. You'd have to have a dual Woodcrest rig to confirm that the bench is correct or incorrect. Almost every time I've complained about a WU being underbenched based on performance on a rig similar to the benchmark machine, I've found myself to be wrong when I tested on a machine identical or nearly so to the benchmark rig.

Posted: **Sat Dec 29, 2007 4:11 am**

Flathead74 wrote:This is very repeatable and agrees with the reports from my teammates, and others.

One item that has not been mentioned: Changes in the WU.

All WUs are benchmarked at Gen 0. As the protein folds, the calculation times inevitability change as the protein changes shape. I have no idea how much the calculation speed can change, but it does change (sometimes in systematic ways, other times in random ways). That's just a fact associated with the benchmarking method of assigning points.

Posted: **Sat Dec 29, 2007 4:13 pm**

Project: 3061 (Run 0, Clone 49, Gen 1)
Avg. Time / Frame : 8mn 25s - 2273.77 ppd

Q6600 @ 3.3GHz
MSI P35 Neo Combo-F (bios ver 1.4)
1GB DDR2 PC6400
XP Pro sp2

I don't have any times for earlier p3061s, run on this machine.

Posted: **Sat Dec 29, 2007 6:12 pm**

2274/3.3/4=172

172*1.1=190

172 PPD/GHz/core is just within the +/- 10% variance that is typically accepted as normal, even without consideration for hardware differences, or points drift.

Yes, it's on the low side, but not significantly. The other WUs on the high side do average things out over time.

Thanks for the report.

Folding Forum

Project 3062 PPD

Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD

Re: Project 3062 PPD