Folding Forum

cine.chris wrote:...My impression was that recent WUs weren't responding consistently and yielding poor results.
Honestly, I just got the feeling my efforts were being wasted by some grad student that couldn't configure their WUs correctly.
As a retired aerospace engineer, I don't like the feeling that my skills, time, $$$, kWHr are being wasted, when all I expected to see was an accurate & consistent metric. if the metric is broken, I begin to question the process...

Let me attempt to describe this in a different manner... the projects in question are specifically created for COVID Moonshot (https://covid.postera.ai/covid).

Think that the current scientific community wants to create a light bulb. However, no human has created the light bulb. So, a group of scientists decided to join together and test various ideas on how to create a light bulb. Of course, there will be ideas that will fail. Those failures will be used as learning and applied in a continuous learning manner to improve their next set of ideas. If they perform enough iterations, they will began to find ideas that fail (and they can avoid them in the inception stage as opposed to the development/test stage), ideas that might be suitable and ideas that are great. The more they do this, the better the next set of ideas are and they will reach the goal of creating the light bulb that the scientific community wants and that it can be shared across the globe without issues.

In science, failures in experiment isn't a setback, instead, it's an opportunity to learn more about something that they don't know. The key is to not give up (https://www.uky.edu/~eushe2/Pajares/OnFailingG.html). That is what the researchers behind Project 1340X are doing. The failures are generating valuable scientific data which they can use. It's only a matter of time before they have sufficient data and can make massive progress

PantherX wrote:In science, failures in experiment isn't a setback, instead, it's an opportunity to learn more about something that they don't know. The key is to not give up (https://www.uky.edu/~eushe2/Pajares/OnFailingG.html). That is what the researchers behind Project 1340X are doing. The failures are generating valuable scientific data which they can use. It's only a matter of time before they have sufficient data and can make massive progress

“I have not failed. I've just found 10,000 ways that won't work.”
~ Thomas Edison

I wonder if you're seeing the core 21 PPD results, and are unhappy with how they reflect to Core 22 results?

MeeLee wrote:I wonder if you're seeing the core 21 PPD results, and are unhappy with how they reflect to Core 22 results?

I haven't run one of the WU's in question here, but just as some input on Core 21.

I've only had 2 Core 21 WU's since I installed HFM. I don't recall any others before I had HFM, and if there were any I had no way to track them. But having heard that Core 22 projects usually deliver a higher PPD by 10-15%, I was shocked when I got the Core 21 project. In my case it was two WU's of the same project. BUT..... they delivered significantly higher PPD than any other WU I had ever run. With a little searching I also found that some beta testers reported low PPD when it was in beta. I'm not sure if PPD was adjusted up or not. But at any rate this very small atom count Core 21 project seems to have disliked many of the more powerful GPU's that people on the beta team have, yet returned great speed and PPD with my onboard graphics.

So I think the complexities of various work units are much greater than we realize, and for any given work unit there might be hardware that struggles while other hardware excels. I personally see a trend that it goes well beyond atom count as well.

The bonus points get complicated. A few years back while Dr Pande was still heading up F@h, he posted a graph that one of the grad students working for him of the total points earned for WUs from various projects versus the power of the GPUs processing them. The curves could only coincide for parts of the distribution over the GPU power, either at the low end or at the high end.

Basically too many variables different between projects. So a high atom count WU will process well on a high end GPU with many shaders, utilizing as many of them as possible. While a low atom count WU utilizes some fraction of the shaders, and spends more time transferring data in and out of the GPU over the PCIe bus.

A low to mid range GPU with fewer shaders on that high atom count WU will just use all of them, and also be doing the same on the low atom count WU. So a WU that does not run so well on a high end GPU may just come out ahead on points on a lesser GPU compared to a WU from an different project. But the high end GPU should still finish quicker.

Probably other factors as you mention besides raw atom count such as the geometry of the protein being simulated in the WU.

As for the 10-15% improvement for Core_22 compared to Core_21, early tests to validate the Core_22 code were run on some projects that were clones of ones that had been done before on Core_21 to see if the same results were obtained. Testers found the WUs completed in less time than when run on Core_21. Improved code, a newer version of OpenMM, or whatever else gives Core_22 a bit better throughput.

BobWilliams757 wrote:In my case it was two WU's of the same project. BUT..... they delivered significantly higher PPD than any other WU I had ever run. With a little searching I also found that some beta testers reported low PPD when it was in beta.

Simple solution: Join the beta team and point out the folly of their "low PPD" reports. Greater diversity makes for more realistic reports.

So I think the complexities of various work units are much greater than we realize, and for any given work unit there might be hardware that struggles while other hardware excels. I personally see a trend that it goes well beyond atom count as well.

I can understand the trends that are dependent on atom counts. I'm not sure beyond that but I'd be interested in your observations.

Nevertheless it's pretty nearly impossible to satisfy everybody though we do try.

Joe H,

Thanks once again for the background story and how it all came together over time. I folded years ago for a period (dial up and no points that I recall), but failed to follow as the project developed over time.



Bruce,

I'm already considering the idea to volunteer for beta testing. At one point I paused due to the fact that so many beta testers have higher end stuff and multiple rigs/GPU's within a rig, etc. Due to that projects move quickly through beta. But in hindsight, maybe having some lower end gear in the mix might help them identify certain strange WU trends or something.

And for my gear, atom count alone does not seem to create the trend. From my (limited) digging, it seems the number of checkpoints and/or steps (if not the same) also factors into the picture. One of the largest atom count project I have run is 448,584 atoms. currently the third largest atom count of any project. But on my system, my PPD was above average and above a number of projects with half or less atom count. In my case the onboard graphics is somewhat unique to most folding systems, so maybe it isn't impacted in the same way as most dedicated GPU's, since memory allocation and such is more variable. Or it could get into the geometry of the protein as Joe H mentioned.

But I think PPD returns are a fairly complex thing overall, and certain hardware will be an outlier no matter how hard people try to find a consistent credit. And the recent PPD and hardware database seems to show that every GPU has strong and weak points, even my low end onboard stuff.



As for your last statement, let me clarify that I'm in no way not satisfied with F@H in any way. High, low, or no credits, if they keep returning OK I'll keep folding them. And I do appreciate you long time folders that have been around guiding the new crowds for years.

I think it is useful for beta testers to have a variety of kit, I recently added a GTX1660ti not only because I found a great open box price but I wondered how it would handle some work units differently being of this strange generation of Turing cards.

Personally, I don't spend a lot of time in the Beta test area but I do have a wide variety of older equipment. At times I do download a WU or two just to confirm my equipment will still run them. When the do (which is 99% of the time) I don't bother with a report because the project owner mostly pays attention to the error reports.