Hi. I am returning to the climate modelling project after several years absence. I have installed the latest version of BOINC onto my main PC which has an Intel 4770K Haswell processor and 8Gb of RAM. This has 4 physical cores and 4 hyperthreaded virtual cores. After much faffing about I finally managed to get the project loaded up and retrieve some work. On past experience I was expecting to pick up 8 models to match the 8 cores on the CPU, but I've only picked up 4. Do the models only run on physical cores? Is there some setting I need to adjust in order to used all the cores and 100% of the CPU power? Perhaps those in the know could enlighten me. Many thanks.
Pete McCann

---

Hi Pete

Welcome back.

It's not a setting, just that you ran a bunch of the "short" models which crashed.
(Model crashed: ATM_DYN : INVALID THETA DETECTED.)
When this happens, BOINC starts to limit the amount of work sent. Once your computer returns a few good results, the limit will increase again.

The "short" models appear to be exploring areas of parameter space that are physical unstable. i.e. those physical conditions are impossible, so the program terminates. Which is what the research is about.

Hi.
Thanks for the explanation. Are all of the models now "short"? In the days of the BBC models they used to take months to run on each computer. Is this now a thing of the past? Do people still back up their models in case they go wrong or is that not needed now? Pete

Hi Pete, there are still longer models, I only back up the hadcm3n models which on my machines at least take several months if the computer is not on 24/7 and over a month even if it is. It is a little while since any of these were released. Also stopping these models to back them up is thought by some to cause them to crash!Certainly they are more prone than some have been to this and I have never had one survive a power failure whereas last week both the models I had running came through a mains failure unscathed!Out of three backed up models of this type I only succeeded in getting one to complete so its a case of paying your money and taking your choice I think.

Hi Dave and welcome back.

The models are shorter than they were several years ago. The 160 year hadcm�s that you are thinking of (that took 3300 hours to run on a 1.2 GHz processor) are gone. The longest model now available is the 40 year hadcm3n. On my 2.6 GHz they take about 20 days running 24/7.

---

For what it's worth -- My CPDN work is no longer backed-up, despite having done so every one to three days for years. Tasks are so short that even every-day backups would be problematic. Salvage a crashed one? Sure, but chances are restoration world bring one or more already finished and uploaded task 'back to life.' The loss on other tasks would probably be greater than loss of one failure (which, unless it is the fifth release of the task, will be reissued). It is possible to do a "single run restore" but it is a risky non-trivial exercise and I no longer attempt it, either.

---

"We have met the enemy and he is us." -- Pogo
Greetings from coastal Washington state, the scenic US Pacific Northwest.

I agree with astro.
We've moved on a long way since those early models. If one crashes, forget it and get another one.
The only exception is if you live in an area prone to constant power cuts/failures. Then you can restore the entire workload and continue. UNLESS you're running the "short" models, hadcm3s. These will probably fail if interrupted.

Lots of different types of models now, with more on the way. And the length varies a lot. As does the size of the zips.
One model that was tested had zips of about 90 megs each.

Ahh the good old days of 160 year model runs! You certainly had a sense of achievement if you managed to get one to the end of its run intact. The models seem to be zipping through on my modern CPU. Hopefully the science is just as valuable as before. On balance I think I can do without the hassle of backing up, in light of there being so little lost if one does fail. Thanks everyone for your input. Pete