This is an example of parallel system that i like: http://librestats.com/2012/10/16/r-at-12000-cores/ . i haven't done anything near this scale, although, that would be a goal to aim for. Although, i am not entirely certain if the above link answers your specific query, because the cores in the link are not housed in what would be considered a single computer. Anton is a super computer unlike the libmpi based solutions that I linked to here.
Additional links:
mpi: http://www.mpi-forum.org/docs/http://r-pbd.org/
Hope this helps.
It was usually applied to at least 1,000 cores in the past plus ability of app to use them. First one was Thinking Machines CM line with 64,000+ simple CPU's.
There are ways to manipulate the potential surface of a system so that your molecule of interest reaches all of available phase space within reasonable wallclock time (read: a few days at most). The keyword is "accelerated MD"; there are several methods. The McCammon group at UCSD has done much work in the field. This doesn't require an Anton in the basement, just a regular PC and a modern graphics card.
The real question is: what does David Shaw, the old sphinx, propose to do with his system? I haven't seen much research published, but the common complaint is that he is very slow publishing.
Beautiful. I just spent two weeks wandering around trying to figure out the name of accelerated MD, and suddenly you've got the boldness to educate HN about it.
There is a fruitful space of combining good priors about the pattern of folding expected out of a given protein with MD simulations. I dream that it revolutionizes how we determine the structure of proteins.
You're right about enhanced sampling methods (of which aMD is just one). The D.E. Shaw lab actually presented detailed comparisons between conventional simulation and techniques like this at this year's Biophysical Society Meeting. I don't see why you couldn't also run accelerated techniques on Anton, or perhaps a related ASIC system though.
Now, you don't have to look too far to see what scientists are doing with this system (https://www.deshawresearch.com/publications.html, http://www.psc.edu/index.php/resources-for-users/computing-r...). From the D.E. Shaw group alone, 7 papers in Science, 7 papers in Nature-family journals, and 13 papers in Proc. Natl. Acad. Sci. The publish slowly because they don't have any obligation to publish. I can't vouch for all of those papers, but it's definitely a step in the right direction for the field.
9 comments
[ 2.9 ms ] story [ 26.0 ms ] threadedit: typos, clarity etc.
https://en.m.wikipedia.org/wiki/Connection_Machine
The real question is: what does David Shaw, the old sphinx, propose to do with his system? I haven't seen much research published, but the common complaint is that he is very slow publishing.
DESRES is a weird outfit on a sunny day.
There is a fruitful space of combining good priors about the pattern of folding expected out of a given protein with MD simulations. I dream that it revolutionizes how we determine the structure of proteins.
Now, you don't have to look too far to see what scientists are doing with this system (https://www.deshawresearch.com/publications.html, http://www.psc.edu/index.php/resources-for-users/computing-r...). From the D.E. Shaw group alone, 7 papers in Science, 7 papers in Nature-family journals, and 13 papers in Proc. Natl. Acad. Sci. The publish slowly because they don't have any obligation to publish. I can't vouch for all of those papers, but it's definitely a step in the right direction for the field.