Nina Singhal Heinrichs and Vijay S. Pande. Journal of Chemical Physics (2007)
SUMMARY. This paper lays out how one can revamp FAH calculations to make them considerably more efficient, perhaps by as much as 1000x reduction in the needed computer time. The basic idea is that we use FAH to build a model of the problem in question (a so-called Markovian state model or MSM) and then use the MSM to predict experimental quantities. When using an MSM to make predictions, the question is usually have we done enough computation to make a sufficiently good (precise) prediction. By calculating the uncertainty (precision) on the fly, we can now send FAH clients to the parts of the problem which are uncertainty limiting. We show that this approach can be considerably more efficiently (1000x) than just running with even sampling. This approach is being incorporated into the FAH server code. One exciting ramification of this work is that while MSM’s were originally formulated as a means to use a large distributed cluster (like Folding@home with 300,000 processors) to try to reproduce what a single, hypothetical machine which is 300,000x faster (which doesn’t exist) could do. However, even if that 300,000x faster machine did exist, we show that our approach would be more efficient than a single, long trajectory, suggesting that MSM-based methods should be useful for a very broad set of computer hardware, not just distributed computing platforms.