What gets the best result? HT on, or off?
HT on results in slower finishing times, but more tasks per run.

On a 4 core 8T Core i7, or a 10C/20T Xeon processor, is there a lot of L-cache swapping going on when using 2 projects per core?
If so, maybe it might benefit running less tasks (disabling HT)?

It depends on the science application.

It depends on the science application.

To add to what Les says, some applications scale so you double the number of threads, you double the throughput. Others it is a law of diminishing returns so from four cores say on a four core 8 thread CPU. each additional core after four will be less of a gain than the one before.

Still others on some 20 core 40 thread machines, lack of level 3 cache results in all tasks crashing.

The different cases above all happen with CPDN. I imagine with other projects it may be the same, certainly with WCG there is big variation in the demands of tasks. Some projects the same rules may apply for all tasks.

I'm not talking about project specific, since my Boinc has a mix of several projects at a time.
Just overall what would be best?
(I'm usually running about 5-10 Einstein projects, a Moo! Wrapper project, some Milkyway projects, and a few others on a 10 core / 20 thread Celeron).

I'm not talking about project specific, since my Boinc has a mix of several projects at a time.

My guess would be to leave it on as with a mix like you have some will always be ones less likely to hog lots of ram/cache but I am willing to be told differently by those who have actually got cpus that support hyperthreading!

The memory isn't that bad.
Rosetta uses up the most, round about 7GB for 20 threads.
Then there are other projects that barely use 35MB on 16 cores!
My system is equipped with 32GB, but I'm thinking of running it with 16GB instead. More than enough memory.
GPU projects use mostly VRAM, and very little ram.

My system is equipped with 32GB, but I'm thinking of running it with 16GB instead. More than enough memory.

Sounds like fine for your mix of projects. I only know in any detail about CPDN. There are some tasks at the moment which run into problems if there is not enough level3 cache rather than main memory. At some point having made it through testing, there are likely to be openIFS tasks that in testing were using over 5GB of memory/task. There are a few other projects use that much, mostly I think ones where you need to run Virtual Box so if running LHC CPDN and possibly one or two others it would pay to check on how things are going from time to time. When I get around to giving this desktop a heart transplant (Well transplant of everything except case and PSU to be honest) I will probably keep a real core free most of the time to be on the safe side.

There is no hard and fast rule for hyper-threading on/off. You just have to try each, for reasonable length of time (days or weeks), across your mix of projects and see which works best for you.