Folding Forum

By weighted, what I'm referring to is the scientific value of the WU.

My understanding of all F@H work units is as follows: you receive a tiny, ridiculously small slice (a WU) of a sector (a project) that is part of a pie (F@H). The impression I get everywhere, on these forums and others, is that you should run a GPU2 client (preferably on nvidia hardware) to maximise PPD. However, AFAIK, GPU's can only simulate protein folding in a specialised way, insofar that each WU, while fast and efficient, is contributing to a much smaller sector of the F@H pie. My understanding also tells me that the unicore 'classic' client has the broadest scope for WU calculation, but is relatively slow. The SMP client, again, is more specialised than the unicore client but is very efficient due to multi-threading on x86 code.

Ultimately what I want to know is: If the scientific importance of a GPU WU, for example, is just as great as a unicore WU, then where is the incentive to crunch on PS3's or unicore clients? I could ask this for GPU vs. SMP, Unicore vs. GPU etc.

Shouldn't the PPD of unicore client projects or PS3 WU's or whatever be increased as to get more people crunching on a wider variety of WU's? I can see how the SMP2 bonus scheme is encouraging fast returns, but why should someone run their PS3 or unicore client when GPU's etc are returning 2, 3 and 4 fold PPD?

Is the science from these clients not as 'important' than those from the GPU client? Or are the projects from SMP and GPU clients really valuable?

The following is just conjecture on my part, I do not speak for the Pande Group but it is my guess as to why units are given the values they are.

The points are weighted in favour of the high-performance (PS3, GPU and SMP) clients over the unicore clients. It is true that these clients are less versatile than the unicore client at the moment but the weighting is to reward the additional speed this gives the project. I have already seen one set of uniprocessor WUs where they were doing the final intensive analysis of a series of projects that had been run on GPUs and PS3s first to narrow down the area that needs studying. In this respect the less flexible but much more speedy High-Performance clients have enabled the Pande Group to reach the final stage much quicker than if the whole thing had had to be run on unicore clients. As speed is everything in this project (so long as the results produced are accurate) the Pande Group sees fit to reward those fast returns with much higher PPD.

I think there is probably a massive amount of machines out there running the unicore client for lack of being able to run anything else. It's also the only client that requires zero babysitting, and is brainless to set up. It may offer the least amount of points, but it's painlessness and it's ability to run on everything out there, will continue to draw many users. Lots of points to encourage use is not required methinks.

John Naylor wrote:The points are weighted in favour of the high-performance (PS3, GPU and SMP) clients over the unicore clients. It is true that these clients are less versatile than the unicore client at the moment but the weighting is to reward the additional speed this gives the project. I have already seen one set of uniprocessor WUs where they were doing the final intensive analysis of a series of projects that had been run on GPUs and PS3s first to narrow down the area that needs studying. In this respect the less flexible but much more speedy High-Performance clients have enabled the Pande Group to reach the final stage much quicker than if the whole thing had had to be run on unicore clients. As speed is everything in this project (so long as the results produced are accurate) the Pande Group sees fit to reward those fast returns with much higher PPD.

Ahh, I see.

So the uniprocessor client does the final work/calcs for a particular WU that has been 'filtered' through the high speed clients.

Thank you Mr. Naylor.

hj47 wrote:

John Naylor wrote:The points are weighted in favour of the high-performance (PS3, GPU and SMP) clients over the unicore clients. It is true that these clients are less versatile than the unicore client at the moment but the weighting is to reward the additional speed this gives the project. I have already seen one set of uniprocessor WUs where they were doing the final intensive analysis of a series of projects that had been run on GPUs and PS3s first to narrow down the area that needs studying. In this respect the less flexible but much more speedy High-Performance clients have enabled the Pande Group to reach the final stage much quicker than if the whole thing had had to be run on unicore clients. As speed is everything in this project (so long as the results produced are accurate) the Pande Group sees fit to reward those fast returns with much higher PPD.

Ahh, I see.

So the uniprocessor client does the final work/calcs for a particular WU that has been 'filtered' through the high speed clients.

Thank you Mr. Naylor.

That is just one example, many uniprocessor projects have not been "filtered" first, but it is one way the high-speed clients have been used. Glad I could help

Not exactly sure what you mean by "weighted" but yes, the benchmark points for each of the client types have been adjusted to reflect their speed and scientific value to the project. The scale is not linear, so I guess one would call that weighted. Some clients are really fast, but can only do a limited type of simulation. For example, a PS3 may be 25x faster than a CPU, but it can only do 1/4 the type of simulations needed by the project.

There is a points FAQ that may help explain some of this. http://folding.stanford.edu/English/FAQ-Points

To answer your questions... A PS3 work unit valued at 900 points has the same scientific value of a CPU work unit valued at 900 points. But note that a PS3 will finish that work unit in a day, and the CPU may take several days. The benchmark value for the PS3 probably reflects that speed.

Your second question... while scientific value of the work unit is the primary concern when setting the benchmark score, or maybe "weight" in your context, the project does take other concerns in to consideration. For example, there was a Big WU bonus for CPU work units that doubled the points on work units that took extra bandwidth to upload/download, or that used a lot of system memory. That is an exaggerated example by comparison to other considerations, but you get the idea.

As to the last question, it's a glass half full answer. Yes, the SMP and GPU work units are really that much more valuable. The CPU work units are not less valuable so to speak. Their PPD has not been reduced, they just earn less by comparison. But again, consider the requirements to run each type of client. The CPU client is a no brainer to install and run. Set it and forget it. Also works well on systems that are slower, or only run part time as the deadlines are much longer. The SMP client is harder to install, requires a small amount of babysitting and updating, requires ~4 times the resources with deadlines ~10 times shorter.

As you become more familiar with the project, and read more of the FAQs, these intricacies reveal themselves.

It's starting to make a lot more sense, thanks 7im