I haven't seen it. It is slightly odd. You wouldn't think that a single point on a climate projection would be that critical, or that hard to weed out if it were bad. But I know nothing of their methodology either.

This raises a LOT of questions concerning the validity of the output of this project. This is output from a weather forecast model (WRF). As such, there could be a deviation from the model output and actual observational data. What that means, is bad data is being fed into the next 48 hour run and so on through the sequence and unless there is a reset back to a set of known observational data the error is going to multiply. The end result is output that doesn't reflect actual conditions in 2 years. Each 48 hour run should start with actual observational data not output from a previous model run. If the goal is to start with a set of observational data and do a forecast two years out then the process is probably valid but I would bet most forecasters would tell you that the two year forecast is not going to reasonably resemble actual conditions that far out. If we could forecast two years out with any kind of accuracy we would be doing it now. Maybe the goal of the project is to see how much the forecast deviates from the actual observed conditions. Most reasonable people would say the deviation is going to be quite large. Is it worth the computational effort to quantify the magnitude?

Has anyone figured out how ARP behaves??? If I run it exclusively on a computer it slows down tremendously even if I have enough RAM (minimum 1 GB per WU).
None of the Xeon E5s or i7s kick the CPU up to its turbo frequency and they just stay stuck on the base frequency. My i9-9960X will increase the CPU frequency a few percent but no where near its turbo frequency.
The i7-6950X can finish an ARP WU 3 hours sooner than and i9-9960X (21 vs 24 hr) but the i9 as a whole completes more work per day than the i7 (32t vs 20t).
Is hyperthreading choking this project? Once I clear the deck I'm going to disable HT on the i9 and see how it runs.
There's something going on that seriously slows this project down but I can't put my finger on what it is. Anybody have a suggestion???

They're all Linux Mint 19.3 (Ubuntu 18.04 Bionic). I'm working on a table to summarize.
The i9-9960X has:
Level 1 cache size = 16 x 32 KB 8-way set associative instruction caches & 16 x 32 KB 8-way set associative data caches
Level 2 cache size = 16 x 1 MB 16-way set associative caches
Level 3 cache size = 22 MB 11-way set associative shared cache

The i7-6950X has:
Level 1 cache size = 10 x 32 KB 8-way set associative instruction caches & 10 x 32 KB 8-way set associative data caches
Level 2 cache size = 10 x 256 KB 8-way set associative caches
Level 3 cache size = 25 MB 20-way set associative shared cache

Both computers have 4x8 GB DDR4 2666 MHz of RAM.

That's to be expected. Even if you set BOINC to half the possible threads with HT on you will observe faster processing, at least I do on my machine with W10. Nothing needed to be hyperthreaded.