I was wondering if anyone could provide guidance for an optimal (or desired) setting for the Target CPU Runtime preference? Having done a lot of work with neural networks I know that this can be an important parameter for finding local minima... you want to spend enough time to try to find a minima but you don't want to waste time on a dead-end. To use the terrain example, I want to try to spend enough time hiking around to find the lowest spot in the area but not waste time if I started out by being dropped off on a mountain top.

Obviously the system speed impacts the amount of work done in a given period of time but in general has anyone found trends like most of the lowest energy predictions come from systems that did 4-hour runs or 8-hour runs or something like that? Does anyone have any recommendations for a setting for this parameter?

Thanks for your thoughts,

Steve

---

I was wondering if anyone could provide guidance for an optimal (or desired) setting for the Target CPU Runtime preference? Having done a lot of work with neural networks I know that this can be an important parameter for finding local minima... you want to spend enough time to try to find a minima but you don't want to waste time on a dead-end. ...

Rosetta wants a number of runs (sometimes called "models" or "decoys") done for each WU, with each run using a different random number seed. Each run will be independant of all other runs. The WU will be sent to a number of computers, and each computer will do as many runs as its runtime preference allows. The Rosetta program ignores the runtime preference during a run. At the end of the run, the program checks the runtime preference to see if there seems to be time for another run. Thus, a longer run time preference increases the number of models calculated, but has no effect on what the program does during a run.

A longer runtime (8+ hours) does help reduce the load on the rosetta servers, as well as reducing the bytes/day your computer has to download.

each Work-Unit that your computer processes will consist of a number of decoys - each one having reached its end-point. The increased run-time just increases the number of decoys produced.

There are advantages to longer run-times though, including a slight increase in productivity as the initialisation process is only run once at the start of a WU, and then there's the reduced demand on the Rosetta servers as there's no difference in size between a 4hr WU and a 24hr WU.

If your computers are reliable and will return results in time then I'd recommend 24hrs.

HTH
Danny

---

beaten by two minutes! didn't think my typing was that slow!

---

So... sak, you're getting too low level on the math. Rosetta will run in increments of "complete models", and when they build the WUs, the math that defines a complete model is build in to the WU. Some have models that take hours, others can crunch a model every 5 minutes. It just depends on what the purpose for the specific WU is, and the type of math behind it.

My general reccomendation is to select a preferred WU runtime that results in me completing each WU in about a day. So, if your machine is on and crunching Rosetta 24hrs a day (like mine), I've got a 24hr WU runtime selected. If your machine is on 12hrs per day, but you run other BOINC projects and Rosetta gets a 50% resource share, I'd tend to suggest a 6hr runtime. It really doesn't make much difference to the project, and it doesn't make ANY difference to the science you're helping with.

Always heed the caution mentioned here and only increase your runtime gradually. I suggest to increase it only one notch per day. Just avoids confusing BOINC and having too much work and perhaps missing deadlines.

---

Add this signature to your EMail:
Running Microsoft's "System Idle Process" will never help cure cancer, AIDS nor Alzheimer's. But running Rosetta@home just might!
https://boinc.bakerlab.org/rosetta/