Have some large-memory machines (64GB and up) here. Anyone recommend a project that would benefit from the extra headroom?

Rosetta uses 3.5GB per task occasionally.

---

Basically most VirtualBox projects as they set up a complete VM per task.
Keith already mentioned Rosetta which is a very good example.
You may also add LHC@home (ATLAS, CMS) to your candidate list.

Depending on the number of CPU cores per computer you may even need more than 64 GB RAM.

I still have no idea when they will be out of testing but OpenIFS tasks from CPDN have taken upwards of 6GB/task. If they do move eventually to that model type for most Linux work, it it will also mean the end of having to install 32bit libraries to run tasks. I would expect there to be a lot of noise once they do make it out of testing but it is probably six months away or more.

Thanks to all for the great ideas. And, no, my post shouldn't be taken as a sign by developers to consume as much RAM as possible. : )

Has anyone tried to measure the overhead incurred by virtualization vis-a-vis memory or other system resources?

To measure the overhead you would need to compare a scientific application running native with the same application running in a VM.
Some examples where you only have a VM app and no native app to compare:
Rosetta
LHC ATLAS/Theory (on Windows)
LHC CMS
QuChemPedIA (not on Linux)

In any case you would have to make yourself familiar with the resource requirements of the app your computer gets.

With Rosetta they don't have the same application in both virtualised and un-virtualised forms which makes the simple comparison very difficult.

??
That's exactly what I meant.
"... examples where you only have a VM app and no native app to compare"

Rosetta, MLC, LHC, Cosmology, Qchempedia, mind modeling....

Also helps if you have more cores (like, some projects like milkyway, can make use of as many cores as you have in a single WU).

Pretty good rundown for RAM requirements here