It looks like a pretty well put together project just based off of what is in the readme. While I don't have any immediately applicable input I could definitely see this being useful in trying to visualize some of the graph theoretic representations of intermediate values within some manifold learning algorithms (and there certainly could be enough data points that this graphviz alone would be non-feasible (though sfdp might be if I had read those docs right)).
From the paper, it is interesting how different the results of sfdp and mars look like.
Thanks, I put a fair bit of effort into cleaning it up!
Mars and sfdp take very different approaches to graph layouts, attempting the embed the graph metric (mars) vs keeping your neighbors close while pushing everyone else away (sfdp). As a result the drawings look very different and emphasize different features of the graph. We're not trying to say that one gives a better drawing than the other, but that here are two techniques that scale and produce different drawings, one of which may give a better visualization of your data. However one clear advantage of mars over sfdp is mars' ability to easily incorporate edge weights.
You did read the documentation right. Sfdp is just as scalable, if not more so, as mars. Although I've never run mars for as long and on as powerful of machines as Yifan has run sfdp.
From the paper, it is interesting how different the results of sfdp and mars look like.