Some Folding performance numbers (updated after 24 days of testing...)

Re:Some Folding performance numbers (Simba123)

Fri, 10 Feb 2012 17:20:43 GMT

I think Jedis Tracker interface makes all the difference. Makes setting it up and running very smooth and easy. Never had any stability issues with it.

V7 on the other hand, I had a lot of stability problems, along with what I believe to be an overly complicated set-up process.
I see no reason why they could not have simply programmed in drop boxes for the 'slot' options instead of having to remember or look up over complicated setup instructions.
For something that was supposed to make folding easy for a non-computer orientated person, in that aspect it is IMHO a complete fail.
that's me though.

I would have thought that the new client would have been more efficient and produce better results, but on a purely ppd analysis that appears not to be the case.
On SMP units anyway. Those GPU units appear to be the same as on the older client.

Re:Some Folding performance numbers (xanderf)

Thu, 09 Feb 2012 19:46:40 GMT

Well, that's definitely something interesting.

Still, that said, the numbers I was running were using the v7 client throughout - so should at least be proportionally correct in relation to each other.

(I'm surprised you had issues with v7, though. I tried starting up Folding several times over the years on v6, but could never get around random crashing and conflicts between the multiple instances needed for SMP and GPU Folding. It was really just a VERY unfriendly application to multiple-user PCs. v7, though...super easy and smooth to run, multi-user PC no problem, etc.)

Re:Some Folding performance numbers (Simba123)

Thu, 09 Feb 2012 19:35:14 GMT

Those projects (7905 & 8001) are specific to the v7 client.
I'm not near my folding machine so I can't check; I only ran v7 for a short period of time before reverting back to Tracker - I found v7 to finicky and for something that was supposed to simply folding, way to difficult to configure and manage - which is the v6 client.

It is interesting to note the significant difference between the ppd of the clients(projects). It's upwards of 20%!

Re:Some Folding performance numbers (xanderf)

Thu, 09 Feb 2012 19:11:30 GMT

Simba123

you should be getting
101xx -20-24k ppd
6xxx - 26-28k ppd
75xx - 28-32 ppd.

Well, looking at my last dozen projects, I've never gotten any of those. (Looks like I'm mostly getting 7905 or 8001)

Don't suppose client matters? I'm running the v7 client.

(Note that the client PPD estimation has consistently, throughout the run, reported about 10-15% higher PPD than I actually see when looking at my daily stats on EOC or Hardfolding)

Re:Some Folding performance numbers (Simba123)

Thu, 09 Feb 2012 18:18:40 GMT

hmm, well I can confidently say you are not getting the most out of your cpu/smp for some reason.

you quote for your system
19,912 ppd Core i7-2600K @ 4.5ghz + 16gb DDR3-1600 CL9 memory - assuming that's 16gb as per your modsrigs

My machine with 8 gigs running -smp 8 (no adv methods) will average 10k ppd more than that.
That ppd has been verified many times over in discussions with other users with the same settings.

I'd be running SMP -8 on its own for a while and if your ppd is still averaging that number, have a look in task manager and see if you can find whats holding you up.

you should be getting
101xx -20-24k ppd
6xxx - 26-28k ppd
75xx - 28-32 ppd.

Re:Some Folding performance numbers (xanderf)

Thu, 09 Feb 2012 15:27:51 GMT

FWIW, this was a machine run 24/7 for Folding...I did other work on it, but I never stopped Folding during the test window.

I believe the PPD difference you were seeing is probably more a result of using the 'advmethods' on the GPU...I was not doing that.

And certainly, on a day-by-day basis, I definitely did see the daily PPD jump all over the place. That's why I wanted to let it run over such a long window of time to really get a feel for, over hundreds of work units, what the true averages would be. You'll note that I even went over 45k PPD...for one day. It's important to average those out...

Re:Some Folding performance numbers (Simba123)

Thu, 09 Feb 2012 15:22:42 GMT

I think this really only applies to machines that are not dedicated folders.
My rig, which is almost exactly the same as yours - Asus P8P-67 Pro i7-2600k @ 4.5 and cas9 ram@ 1600Mhz (8gig) produces very different numbers

CPU only will average 28k ppd
CPU + 460 stock (GPU3 units) 32k ppd
CPU + 560 ti @ 880 (adv-methods gpu3) 40-42kppd

all 3 together average 46k

I have found running adv-methods GPU3 units results in quite an increase in SMP ppd (presumably because they require less frequent CPU cycles)

Also depends on how many 101xx units you get on your SMP. those units run anywhere from 8-12k ppd less than other SMP units.

I have come to the conclusion that the only way to justify folding SMP and GPU simultaneously is to have them running 24 hours. Otherwise you can get basically the same ppd from running SMP alone because it will run faster and get more bonus points.

I guess each rig has its own peculiarities and usage, so you have to work out what nets the best ppd on your system. As Xanderf has pointed out, it can/does take long runs of many days, and combinations of hardware usage and settings before you find your sweet spot.

It's a bit like old-school FSB over-clocking. You can easily spend weeks trying to squeeze out that last hundred ppd!

(it's almost as addicting too. I currently running a long set of combined SMP + 2 GPU, fiddling with the gpu clocks to try and find the balance between higher gpu clocks/points and smp points, makes for a BIG excel chart)

Re:Some Folding performance numbers (xanderf)

Thu, 09 Feb 2012 12:01:00 GMT

It's really only in the longer-running tests that this variation consistently shows up. In the PPD estimates any client generates on its current job, or even day-to-day numbers, don't really show it that well.

By doing the longer 6-day runs, it definitely did a better job highlighting the result differences.

Re:Some Folding performance numbers (esCob4r)

Thu, 09 Feb 2012 03:20:12 GMT

Nice find! I remember asking if folding was affecting by memory at al and I was told no. Looks like these tests clearly show otherwise. +1..

Thank you for taking the time to run these tests. BR?

Re:Some Folding performance numbers (xanderf)

Wed, 08 Feb 2012 22:53:21 GMT

Another week slogged through, and I've updated the second entry with the new set of data.

At a high level, then, we have after 24 days of testing:
PPD       Configuration
19,912    Core i7-2600K @ 4.5ghz + 16gb DDR3-1600 CL9 memory
28,738    Core i7-2600K @ 4.5ghz + 16gb DDR3-1600 CL9 memory + eVGA GeForce 560 Ti SC (stock)
23,243    Core i7-2600K @ 4.5ghz + 16gb DDR3-2133 CL9 memory
31,737    Core i7-2600K @ 4.5ghz + 16gb DDR3-2133 CL9 memory + eVGA GeForce 560 Ti SC (stock)

...in other words, from a 'baseline' Core i7-2600K @ 4.5ghz with DDR3-1600 memory (not that unusual a configuration), we see:

16.7% increase by bumping memory up to DDR3-2133 over DDR3-1600 with CPU on its own
10.4% increase in PPD by memory adjustment with GeForce 560 Ti in the system
~8800 PPD increase by adding a GeForce 560 Ti to the base system with slower ram
~8500 PPD increase by adding a GeForce 560 Ti to the base system with faster ram

The long and short of it is that ram speed makes less of a difference if you are using GPU Folding (although it still makes a difference), but if you are planning on going CPU-only...you definitely want that fast ram. Presumably, the reason for this is that the GPU Folding rate is virtually unaffected by system ram speed...the GPU Folding providing basically a flat baseline of PPD. However, when using GPU Folding with CPU Folding, the GPU siphons off performance from the CPU resulting in smaller variances in CPU-related Folding performance.

Re:Some Folding performance numbers (xanderf)

Thu, 02 Feb 2012 22:39:13 GMT

I've updated the above post to indicate the memory performance speeds I was using.

To test the impact of memory performance, I reduced my system memory speed (changing no other variables) to DDR3-1600 from DDR3-2133. Same memory timings in every degree.

Core i7-2600K @ 4.5ghz + 16gb DDR3-1600 CL9 memory + eVGA GeForce 560 Ti SC (stock)
DATE          PPD         Work Units
02.08.12      26,741      16
02.07.12      34,601      17
02.06.12      26,113      16
02.05.12      29,532      13
02.04.12      29,248      13
02.03.12      26,195      22

Average PPD: 28,738

Core i7-2600K @ 4.5ghz + 16gb DDR3-1600 CL9 memory, only
DATE          PPD         Work Units
02.02.12      20,579    23
02.01.12      19,610    6
01.31.12      20,048    13
01.30.12      21,132    12
01.29.12      28,500    16
01.28.12      9,602    2

Average PPD: 19,912

...the oddness of the work units on the first two days is basically just due to the breakdown of when a job got submitted vs the midnight boundary, and the reason I use large test runs of 6 days. Note that the average of even just those two days is very close to the same as the overall average for all 6.

Some Folding performance numbers (updated after 24 days of testing...) (xanderf)

Thu, 26 Jan 2012 00:07:12 GMT

So I've been running my system over the past 2 weeks with two very distinct configurations, to get some data on *actual* PPD results.

I'm planning on some more tweaks, so should be updating this as time goes by. To start with...

Core i7-2600K @ 4.5ghz + 16gb DDR3-2133 CL9 memory + eVGA GeForce 560 Ti SC (stock)
DATE          PPD         Work Units
01.25.12      28,874      32
01.24.12      45,223      19
01.23.12      25,984      27
01.22.12      31,020      27
01.21.12      26,244      30
01.20.12      33,078      26

Average PPD: 31,737

Core i7-2600K @ 4.5ghz + 16gb DDR3-2133 CL9 memory, only
DATE          PPD         Work Units
01.19.12      23,306      5
01.18.12      25,730      5
01.17.12      25,948      5
01.16.12      19,108      4
01.15.12      21,758      4
01.14.12      23,607      9

Average PPD: 23,243

This is not a dedicated Folding machine, just a PC that will be experiencing 'average day to day workloads'. And, obviously, I'm not controlling for the specific projects outside of simply including as large a data set as possible (32 work units across 5 days even for just the CPU test) to trend out the effects.