Projects 10012-10085 -> accelerate F@H 100 times?

[Jonazz]

I've recently read the project description for projects 10012-10085, and I was shocked by what I read, especially this part:

This project is a test of a novel means to significantly accelearate Folding@home Molecular Dynamics calculations by 100x without any additional increase in hardware

WOW! How realistic is this? And will F@H literally run 100 times faster or does it mean something else?
If this is serious, why not devote more systems (GPU, SMP, Bigadv) to it? F@H's potential would be huge!

[7im]

The test is real. Don't know the results yet.

100x the speed with with the same accuracy would be great. 100x the speed with 2x less accuracy might be okay. 100x the speed with 10x less accuracy might not be accpetable. Have to wait and see.

No need to devote (waste?) more systems without knowing the test results.

[Jonazz]

The test is real. Don't know the results yet.

100x the speed with with the same accuracy would be great. 100x the speed with 2x less accuracy might be okay. 100x the speed with 10x less accuracy might not be accpetable. Have to wait and see.

No need to devote (waste?) more systems without knowing the test results.

What I meant was, why not devote other systems to get the results quicker? Although I do understand that every researcher is competing for resources, it does seem to me that this might have higher priority.

If accuracy were to be wasted however, I don't think that's okay. We already have the GPU's for that

[7im]

...
What I meant was, why not devote other systems to get the results quicker?

That is a good question. And you gave one good possible answer. I don't know any other possible answers this late at night, but your on the right track.

[FlipBack]

Wow this is exciting! Can't wait to see results.

[GreyWhiskers]

Just picked up my first 10085 on my old HP Uniprocessor.

What's on the ProjSummary is considerably "richer" in points than was described a couple of weeks ago.

The points went from 242 to 600
The timeout went from 4 days to 8 days
the deadline went from 8 days to 11 days
AND, a QRB k-factor of 0.75 is in effect.

Big contrast to the p3866s I was getting for the last few weeks - only two of which ever successfully went to completion.

Code: Select all

10085	129.74.85.15	ww22	569	8.00	11.00	600.00	100	GRO-A4	Description	abdulwahidc	0.75

Projects 10083 - 10085: adv -> full
by abdulwahidc » Thu Jun 16, 2011 5:04 pm

Projects 10083 - 10085 moved to full

These use the GRO-A4 core on win32. The simulations are of several WW mutant structures (the wild type, slowest folder, and fast Fip35 mutant) found in Liu et al. [1] running at different temperatures.

242 Points
4 day Timeout
8 day Deadline


[1]: Liu F, Du D, Fuller A a, et al. An experimental survey of the transition between two-state and downhill protein folding scenarios. Proceedings of the National Academy of Sciences of the United States of America. 2008;105(7):2369-74. Available at: http://www.pubmedcentral.nih.gov/articl ... e=abstract

[izaguirr]

All of our A4 projects on 129.74.85.15 were lacking k-factors. We have adjusted this recently for all A4 projects to be more in line with the point guidelines by the Pande Group. That's why you see the changes. Thank you

[izaguirr]

@7im: yes, you've summarized our approach well.

For those of you who are curious, by the end of the summer 2011 we hope to have characterized the performance / accuracy tradeoff of this methodology. Things are looking good. Then we hope to deploy it for select projects on an OpenMM GPU core by the end of 2011. That means that the methodology will be at first restricted to the types of simulations that can use GPUs.

A second stage would include extending the methodology to work with general simulations that can be then incorporated into core A4 (or its successor). We plan on doing this in 2012.

Even at its most successful though, different scientific projects require different levels of modeling fidelity/accuracy. Thus, while we hope that this methodology can enable projects spanning much longer scientific timescales and eventually much larger systems, don't expect the standard CPU and GPU cores to go away ever.

[FlipBack]

Thanks for the update!

[Jonazz]

Any updates on this subject? I noticed that somme projects have been added (10086-10090) but the older ones are still active as well.

[Stonecold]

100x the speed with with the same accuracy would be great. 100x the speed with 2x less accuracy might be okay. 100x the speed with 10x less accuracy might not be accpetable. Have to wait and see.

100x the speed with 10x less accuracy would be fine. All they'd have to do is give the same WU to 10 computers to ensure the overall accuracy doesn't change. If they all work at 100x speed and each WU is processed by 10 machines, then it would amount to a net increase of 10x speed of overall progress. So that means that even if the accuracy was reduced 50x and the speed was increased 100x, then there would still be a double in the speed of the overall progress (assuming each WU was sent to 50 machines to ensure accuracy).

[Jonazz]

100x the speed with with the same accuracy would be great. 100x the speed with 2x less accuracy might be okay. 100x the speed with 10x less accuracy might not be accpetable. Have to wait and see.

100x the speed with 10x less accuracy would be fine. All they'd have to do is give the same WU to 10 computers to ensure the overall accuracy doesn't change. If they all work at 100x speed and each WU is processed by 10 machines, then it would amount to a net increase of 10x speed of overall progress. So that means that even if the accuracy was reduced 50x and the speed was increased 100x, then there would still be a double in the speed of the overall progress (assuming each WU was sent to 50 machines to ensure accuracy).

But that's not how FAH works. Each WU is supposed to be processed only once.

[Stonecold]

I'm saying IF it were necessary to increase accuracy reduced by a newer version of the client, then that would be a viable solution.

[ChasR]

Would it? All 50 WUs could be returned with the same coordinates and still be wrong if the simulation process is flawed. That has happened before with GPU1 and with QMD work units.

[7im]

izaguirr said I characterized the situation well, so not much need to speculate. Pande Group knows how to balance speed and accuracy better than any of us, so I'm happy to let them decide how best to proceed.