10
$\begingroup$

The Anton computers are purpose-built by DE Shaw research for molecular dynamics simulations. They achieve performance up to an order of magnitude faster than can be achieved with commodity hardware, but don't seem to be used all that much by the wider computational chemistry community. This is despite the fact that limitations on simulation timescales is a major impediment to using simulations to guide experiments, leading to all kinds of intricate theoretical methods to deduce longer-timescale properties from relatively short simulation trajectories.

Having read a bit about the Anton machines, which have been around for over a decade now, I'm perplexed as to why they haven't been more widely used. Is it because they are too inflexible, not available to enough researchers, not helpful in solving scientific problems, or is there some other reason I'm missing?

Improve this question
$\endgroup$
8
  • 4
    $\begingroup$ Perhaps 10 times as much commodity hardware is cheaper? $\endgroup$ – Jon Custer Dec 9 '19 at 0:07
  • 2
    $\begingroup$ And does DE Shaw even make the purpose-built chips available for purchase? $\endgroup$ – theorist Dec 9 '19 at 4:02
  • 2
    $\begingroup$ @theorist No, not as far as I'm aware, but there is at least one machine available for researchers to request time on: psc.edu/anton-rfp $\endgroup$ – mszep Dec 9 '19 at 7:35
  • $\begingroup$ Do you have evidence that the computers are not widely used? Can you quote some numbers? $\endgroup$ – Buck Thorn♦ Dec 9 '19 at 12:07
  • 3
    $\begingroup$ According to this 2016 article, "As far as anyone is aware, there is only one Anton machine outside of the company’s New York-based research lab—at the Pittsburgh Supercomputer Center (PSC)" (nextplatform.com/2016/02/04/…). I also gather from the article that DESRES has kept the technology needed to make it close to the vest, so that's one key barrier. A second possible barrier is cost, and we can't assess cost if DESRES has never sold the tech to anyone else (it was donated to PSC). $\endgroup$ – theorist Dec 9 '19 at 22:16
7
$\begingroup$

Protein dynamics occur in multiple different time scales. I myself is a person using MD simulation for my research. One reason I am not so much interested in ANTON is not it's cost, limited resource etc. It's rather because one extremely long simulation wouldn't help much at least for my research.

Protein's thermodynamics as well as kinetics all depend on free energy landscape which is a priori unknown. To directly compute free energy from simulation, you generally need multiple order of magnitude faster machine even than ANTON. However, there has been many techniques such as Umbrella Sampling that helps resolve this issue on computers in our lab.

I would appreciate it if I can run my Umbrella Sampling in ANTON but it would be so much waste of time to wait on the line if I can run the same simulation for a week in my workstation.

Another reason I can guess why ANTON has not been widely used is because of its specificity. ANTON does not run typical MD simulation packages such as GROMACS, NAMD, AMBER etc. All packages use same algorithm to run the simulation but the commands, file structures etc, are all different. Typical MD user would prefer to use package that he/she is familiar with. I am not sure how much I would have to study to get used to ANTON MD package but why bother if I am not desperate?

So in summary, only those users who are interested in actual trajectories of protein and water in millisecond time scale would want to use ANTON but not many others. Other than such purpose, there are many techniques and algorithms that would work better than running a single extremely long simulation.

Improve this answer
$\endgroup$
2
  • 1
    $\begingroup$ Thank you very much for this answer, it's exactly the type of information I was looking for. Is Free Energy an experimentally measured quantity? Also, if I understand correctly, Umbrella sampling does not involve solving the equations of motion, but using a clever sampling trick to make sure the whole coordinate space is adequately covered, is that right? $\endgroup$ – mszep Jan 2 at 17:55
  • 1
    $\begingroup$ Well, is free energy an experimentally measured quantity? I am not an expert on free energy measurement so I would advise you to google about it for more details. About Umbrella Sampling, it is quite the opposite. US DOES solve equation of motions. In general, it does so with some constraints to sample restricted coordinate space. You run multiple US over space of interest to obtain good samples on the space and combine those multiple data set properly. I mean properly, as properly re-weighting. There are bunch of papers and articles about US, so maybe you want to look it up in the web. $\endgroup$ – Jaehoon Yang Jan 22 at 17:02
4
$\begingroup$

Anton and its successor Anton2 are highly specialized pieces of machinery. In fact, the computer, the chip, and the only code that can run on the chip are all called Anton. There's a reason why the only Anton computer outside of D.E. Shaw is located in the Pittsburgh Supercomputing Center (Anton now deprecated there, replaced by Anton2): this highly specialized equipment can't be used for calculations other than biological MD simulations. Combine this with the fact that Anton is at the cutting edge (and therefore well beyond what most supercomputing centers would deem the threshold for diminishing returns), and you have your answer.

Note: I am in no way bashing Anton. I had the pleasure of running on both Anton and Anton2. While they made huge strides in user-friendliness with Anton2, the reasoning stays the same: Anton was made for MD simulations and nothing else.

Improve this answer
$\endgroup$
2
  • 2
    $\begingroup$ Thanks for your answer, but my question was more about why more groups don't use the Anton machines for their MD work. I suspect that if there had been a long waiting list for access to the PSC machine, Shaw would have made more of them available, and potentially set up a commercial operation? Or am I incorrect in that assumption? Reading between the lines of your response, is Anton [perceived as] too difficult to work with compared to Linux-based supercomputers? Or perhaps the simulation timescale increases are not enough to justify the switching cost? $\endgroup$ – mszep Mar 11 '20 at 13:11
  • 1
    $\begingroup$ There is high demand for Anton. However, given Anton's prohibitive cost and difficulty to maintain and fine-tune (and perhaps D.E. Shaw's hesitation to make it a widespread computer), the majority of supercomputing centers wouldn't be able to justify the cost of owning a machine such as Anton, given that general purpose supercomputers are usually sufficient for MD users (and can serve the needs of other users of the center, unlike Anton). Without being on the board, that's my best guess why the individual group demands for Anton do not translate into mass implementation. $\endgroup$ – jezzo Mar 11 '20 at 14:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.