Explicit/Implicit ratio?

[PlayLoud]

I have watched as CPU TFLOPs have decreased a bit (especially in Linux), while GPU2 TFLOPs have gone through the roof. Since it has been said that both Explicit and Implicit solvent models are needed, I was wondering if there is any sort of ratio that is optimal for the science? Would there ever be a point where too many people were running GPU2/PS3, and not enough running CPU clients? So many people who are building folding systems are going with GPU2 rigs, that CPU folding has taken a hit. Before GPU2, most people building dedicated rigs were running SMP on quad core CPUs.

[Flathead74]

I have watched as CPU TFLOPs have decreased a bit (especially in Linux), while GPU2 TFLOPs have gone through the roof. Since it has been said that both Explicit and Implicit solvent models are needed, I was wondering if there is any sort of ratio that is optimal for the science? Would there ever be a point where too many people were running GPU2/PS3, and not enough running CPU clients? So many people who are building folding systems are going with GPU2 rigs, that CPU folding has taken a hit. Before GPU2, most people building dedicated rigs were running SMP on quad core CPUs.

All that PG need do is readjust the points to shift interest in the direction that they so desire.
It is but a part of the psychological aspect of controlling what gets worked on first,
the social engineering side of the project.

[codysluder]

My guess is that this isn't a big problem mostly because of the huge amount of work there is to be done to fully understand the folding process. Using completely arbitrary numbers (since I don't have any idea what they are) suppose that there are enough clients to do 10% of the explicit work and there are proportionally twice as many flops available for implicit work, we would still be only finishing 20% of the implicit work that the Pane Group wants to do.

They won't need social engineering for a long long time.

[Flathead74]

They won't need social engineering for a long long time.

They are and have already been using it.
That's what bonus points are all about.
It's a way to manipulate what WUs get the main focus of attention.

One example would be the short run of advmethods WUs , 24xx I think, that had were run not too long ago.
These had a bonus added to them so that they would be completed quickly.
It was stated as much when the project was released.
http://folding.typepad.com/news/2008/04 ... for-a.html

[MtM]

I'm with you flathead, absolutely correct

On the other hand, there is an existing unbalance between ppd and scientific value which has been talked about extensively both on the forums and I imagine internally in the PG, yet no attempts have been made to correct the issue. Social engineering with a userbase such as ours is pretty hard because you have people who only care about ppd, people who care about both ppd as well as scientific value not necessarally in that order, and people who only care about the scientific value. Politics are a part of this project now, though I think and hope it's still and thank god for that, a smaller part then the science I hope it stays that way.

[codysluder]

I don't think that everything you're calling social engineering falls into the same category. The original reason for many of the bonus decisions are more closely related to encouraging people to stick with a new development even when it's less than stable. In other cases, it may be that one protein is more important than some other protein.

People should have every right not to participate with the latest (perhaps unstable) technology but they shouldn't be griping about the points that they might be earning if the did. Similarly, those who are earning a bonus shouldn't be griping about a less-than-perfect client as long as it's still being developed.

The bonus for Amber, on the other hand, was based on a clear statement that those projects (for the classic client) were higher priority. Now that the deadline has passed, the priority no longer applies. I don't think the bonus has been removed, as it should have been. If the bonus is still in effect, do you think it should be removed for the projects that are still active? What kind of uproar do you expect that to cause?

The Pande Group does need a method of directing in which way the research is going. Points seems to be one method, but there are other possibilities. One alternative is for the Pande Group to develop tools to force people to run the higher priority tasks and remove the elements of choice/chance. i.e. Prevent people from choosing to run a lower priority protein that may be more nearly optimum for your hardware. Another development (already planned) is to pick projects that run well on your hardware whenever the project priorites are equal. (Of course "run well" is generally measured in PPD which isn't an objective measurement unless all bonus points are removed first.)

Is that what you recommend?

[folding4u]

My guess is that this isn't a big problem mostly because of the huge amount of work there is to be done to fully understand the folding process. Using completely arbitrary numbers (since I don't have any idea what they are) suppose that there are enough clients to do 10% of the explicit work and there are proportionally twice as many flops available for implicit work, we would still be only finishing 20% of the implicit work that the Pane Group wants to do.

If I'm correct, PS3 offers some explicit processing as well so perhaps there's 300-500 teraflops for that. Is that enough so that the speed of research is not curtailed? This is what I'm interested in and I don't think what you said touched on that.

[bruce]

Is that enough so that the speed of research is not curtailed? This is what I'm interested in and I don't think what you said touched on that.

I think it is fair to say that FAH can use a lot more processing power than it has. That will continue to be true for a long, long time. Growth that happens smoothly over time is always better than abrupt changes (either up on down) because there needs to be corresponding changes in server capacity. An excess or shortage of servers (or WUs) for one client cannot be shifted to/from another client without a lot of back-room activity.

[Flathead74]

...The original reason for many of the bonus decisions are more closely related to encouraging people to stick with a new development even when it's less than stable...

This falls directly under one definition of social engineering in that it refers to efforts to influence behavior.