How long does a project run?

[Mactin]

Earlier this week (3rd jan), I successfully folded p5770-r12-c332-g4108 !

Wow ! Gen 4108 !

This means I was the 4109th participant to fold this chain of WUs.

With my GTS250, it takes about 80 minutes to fold this type of WU and postulating the the GTS250 is now kinda average, I conclude that there has been about 5500 hours of processing, ie 228 days 24/7.

To satisfy my curiosity could a MOD indicate when p5770-r12-c332-g0 was folded ?
and to further satisfy it check every 500 gens ?
I would like to get a grasp of WU turnaround (if this is not confidential information).

Thanks

[bruce]

I wasn't able to find any information on the first dozen Gens. Maybe they were completed prior to releasing the Project outside of the Pande Group's test lab. I'm sure you really don't care if I start from the first one I actually found. Here's when each 500 Gens was completed:

2010-06-18 Gen 14
2010-08-27 Gen 514
2010-11-25 Gen 1014
2011-01-22 Gen 1514
2011-03-31 Gen 2014
2011-05-31 Gen 2514
2011-07-27 Gen 3014
2011-10-03 Gen 3514
2011-12-20 Gen 4014

Don't forget that WUs do get lost and reassigned and servers may not be 24x7 every day over 18 months. Moreover, remember that if donors are going to always be able to get a WU, no matter how the number of active clients varies, there has to be a backlog of WUs that are ready to be assigned. In other words, a trajectory cannot run 24x7 because between assignments, it sits on the server for a bit waiting for it's turn to be assigned.

Long term average: 7 Gens per day for this project.

[Jonazz]

I've been getting lots of these and similar projects lately, I guess the GPU2 projects are running dry. But I had no idea GPU projects could run for over 1.5 year, aren't they supposed to be completed much faster?

[bruce]

Most people don't appreciate the magnitude of the problems that are being solved by FAH.

Consider that 1.5 years is much faster than, say, 10 years. That's the reason why there's a QRB that gives you extra points when your assignments are completed faster. Nobody wants to spend a lot of time doing research on a "simple" question that can't be answered within their lifetime, but sometimes that's the only possibility.

(QRB currently applies to all SMP projects and some Uniprocessor projects, with plans to extend it to additional projects -- probably including GPU projects. I wonder how much change, if any, this project would have seen if the QRB had been applied to GPU projects. My guess is: Very little.)

[Leonardo]

Mactin, what a great question!

Bruce, thanks for addressing the question with data. This short little thread is one of the most interesting I've read. That really does drive home what we frequently read, that Pande Group wants the completed work units back as fast as possible. I understood that theoretically, but the data here demonstrates in cold, hard, numbers.

[Jonazz]

But that must means these are some huge projects. Let's say this project takes 2 years to complete and GPU's are 20x faster on average, that would mean that it would take 40 years to run these simulations using CPU's only (not taking speed and core increases in account). That barely makes CPU results a shadow of these calculations. I know the GPU can't run compex proteins with lots of atoms, but I'm not talking about that.
That means that the simulations currently being done with GPU's cant be done on more complex proteins because it would take way too long.

CPu's -> complex protiens, takes 'ages' to process. Or they only need certain parts of a folding protein (those which are relevant).
GPU''s -> 'simple' proteins, but far more data in less time.

If I understand it right (which I probably won't ) GPU's will be far more important in the future because the methods are getting more and more accurate which will allow the PG to tackcle complexer WU's with GPU's. CPU power will only incresae slighty with rising specs and more cores in a single maching (not taking increased number of folders in account).

[Jesse_V]

To clarify, GPUs and CPUs are two very different types of hardware. GPUs use specialized hardware to do specific types of calculations really, really fast, while CPUs are more general purpose and accomplish their calculations at a normal rate. For F@h, GPUs perform specific types of molecular simulations called implicit solvation, while CPUs perform explicit solvation models with atom-by-atom simulation detail. PS3s take the middle ground between generality and speed. By taking advantage of all these different types of hardware, the PG can custom-tailor their software to the hardware, thereby achieving maximum performance overall. As technology improves, hardware will make it possible for them to address problems that were previously considered impossible to tackle computationally.

[Jonazz]

So what exactly is the difference between implicit and exmplicit solvent besides implicit being proteins with les atoms?

It also still doens't answer my questionthat, as time passes by and GPU cores become better and better; won't CPU cores will become less and less relevant?

[Jesse_V]

My understanding is that they are two different models. Explicit is all-atom, meaning every atom, all the surrounding water atoms, all the forces, all right there. Implicit is a little trickier, and I think they try to treat certain things as a mathematical continuum, group atoms together and treat them as one, things like that. But I'm no expert in implicit solvation models, I'm just reiterating the bits that I've read about it. Think of it like modeling an ocean wave. You could either model all the water molecules, or use the sine function. Either way, you get different data from both models, even though the overall thing your modeling is the same size. To be clear, I'm pretty sure that GPUs and CPUs can model the same sized proteins, they just do so differently.

As I said, CPUs have their place, because they do more general calculations which GPUs can't and probably won't in the future, since that's not their job. Isn't CPU technology advancing as well? I just don't see CPUs becoming irrelevant.

[Jonazz]

I think I understand your explanation. And yes, CPU technology is obviosuly advancing as well. But GPU technology is advancing AND the PG's methods of doing folding work with GPU's is advancing as well (not saying they aren't using better and better models with CPU's, but I don't think the increase is as fast), which would make the GPU's more and more relevant as time passes by. That's just my take on things.

[Napoleon]

To put JV's explanation in layman's terms, I vaguely recall 7im making a sculpture analogy somewhere - GPU is like a chainsaw, it gets the rough outline done really fast, CPU is the sandpaper for fine details. PS3... a carving knife, maybe? The point being, PG really needs all these tools to create a polished final product.

[Mactin]

Consider that 1.5 years is much faster than, say, 10 years.

1.5 years for this particular trajectory.
Consider 2.5+ years. I've had these 353 point type WUs (p5765-p5772) running since the very first day I had my GTS250 (1 july 2009) and it was all-ready quite far along (p5771-r7-c114-g763).

I hope that this researcher has not been waiting for his/her data for all that time.

BTW, Bruce, I have not thanked you for answering my original question. Merci beaucoup.

[bruce]

I'm changing the title of this topic and moving it to the science forum. We may get some additional comments in that forum and it's certainly more general that just this one particular example.

[bruce]

My understanding is that they are two different models. Explicit is all-atom, meaning every atom, all the surrounding water atoms, all the forces, all right there. Implicit is a little trickier, and I think they try to treat certain things as a mathematical continuum, group atoms together and treat them as one, things like that. But I'm no expert in implicit solvation models, I'm just reiterating the bits that I've read about it. Think of it like modeling an ocean wave. You could either model all the water molecules, or use the sine function. Either way, you get different data from both models, even though the overall thing your modeling is the same size. To be clear, I'm pretty sure that GPUs and CPUs can model the same sized proteins, they just do so differently.

As I said, CPUs have their place, because they do more general calculations which GPUs can't and probably won't in the future, since that's not their job. Isn't CPU technology advancing as well? I just don't see CPUs becoming irrelevant.

Right. In the implicit model, there is a single equation describing how a uniform field of solvent (usually water) interacts with the outer atoms of the protein. That is much simpler (and faster) than simulating a large enough number of hydrogen and oxygen atoms to adequately describe that interaction. . . . however, the H and O atoms do not create a perfectly uniform field. Water is polarized, in that the H atoms are positive and the O atoms are negative so the electrostatic forces on the protein atoms can be different, depending on the orientation of the water molecule. In some cases, this is important and in other cases it's small enough to be neglected. It's up to the scientists to decide how important those details are to their studies.