I'm a PhD student in algorithmics / algorithm engineering, so I work with a lot of people who do stuff like this, even though my research isn't related to sorting. Super Scalar Sample Sort (which ips4o is based on) was co-authored by my advisor, Peter Sanders, and I re-implemented it a few years ago in modern C++ just for the fun of it. Turns out that was a lot nicer to read than the original code (which isn't public) and a bit faster, too. I put that on GitHub where the blockquicksort authors found it and contacted me and we traded some emails and made some improvements. Sometime later my advisor and two colleagues came up with an idea for in-place super-scalar sample sort, out of which eventually ips4o was born.

So, uh, I guess I just work with the right people :) Sorry that I can't give you anything concrete. All of these papers were presented at ESA (European Symposium on Algorithms), though, so that's a good venue to follow. But beware, ESA has a theory track that's a lot bigger than the experimental track, and papers published there can be somewhat unapproachable ;)

For people who aren't Ph.D. students, it may help to contact a local CS department and see what they have going on. There should at least be a grad student colloquium meeting once a week or so. I attended that at Penn a few times, although it was too far above me to really follow (lots of PL theory). More recently I've been going to a Database Reading Group at Portland State University (http://datalab.cs.pdx.edu/dbrg/index.php), which is a great way to get a taste of current research in that niche. Databases are less rarefied, too. :-)

Thanks, I totally forgot about pdqsort! If I recall correctly, it's partially based on blockquicksort, but includes a lot of other improvements. It's a very interesting algorithm, much faster than introsort etc. However, I believe that IPS4o is still faster overall (I think I asked Sascha, one of the authors of IPS4o, about this a while ago, but I don't have any measurements at hand), even before accounting for IPS4o's parallel implementation. You should definitely check out both :)

It all depends on your machine model. You're right in that the RAM model doesn't model the performance characteristics of real-world computers very precisely. That's why it's a model :) Other models exist that can be used to analyse various aspects of algorithms, e.g. the External Memory model (which can be applied to any level of the memory hierarchy, e.g. to quantify data transfer between cache and RAM). Or you can count the number of hard-to-predict branches, etc.

We almost always assume that a machine word is large enough to describe the input size n. This usually implies constant-time operations on log(n) bits. Not doing so would clutter up notation and complicate analysis, and the whole point of using models is avoiding that where reasonably possible.

I mean you could take this thinking to the extreme and say that as we live in three-dimensional space, the wire length to access more and more memory has to grow at least with the cube root of the size of the memory, because that memory needs to physically be stored somewhere at the end of the day. That's not helpful for analysing algorithms, though :)