Why is it that I get very low credit for a task to takes more than 100,000 CPU seconds? For example, I can complete a task in say 99,000 seconds and get 1,500 credits but a task that runs 100,001 seconds only pays out 300 credits.

sarah

The "problems with..." thread has reported some issues with work units that begin with b21 and b22. On occasion, they are encountering long-running models. Looks like your task hit a long-running model which caused the work unit to run longer than your runtime preference and then be ended by the "watch dog" 4 hours later.

When this happens, your runtime per completed model gets to be extremely long. Since credit is awarded per completed model, you are granted low credit. However, others have run in to the same situation and reported that these models can be difficult, and so the awarded credit per model is slightly higher than it would otherwise be as well. So, luck of the draw, you hit a model that ran poorly, and received poor credit. When you crunch a similar task that runs well, you end up with slightly more credit than you would normally get for the work expended because the history is showing these long-running ones in the average.

The development team is looking in to the issues with these tasks to help them run more consistently.

---

Rosetta Moderator: Mod.Sense

I may have been very unlucky then because I have had many of them go over 100,000 and every single one of them have paid poorly. Does the 100,000 indicate how long the project actually ran or how long it was expected to run? Also, I had my time set for 24 hours.

Right, your 24 hours runtime preference is 86,400 seconds. The "watchdog" will wrap up anything that takes 4 hours longer than your preference, which is another 60*60*4 = 14,400 seconds. So the total comes over 100,000. The figure shown on the website in your tasks list is actual CPU seconds for the task.

It all depends on when in the course of running the task it encounters the problem model. Sometimes 100s of models are done before one that has problems crops up. So, I'd imagine you have a wide variance in credit for these that took over 100,000 seconds.

The project needs to expose failures like these so they can be fixed. So R@h makes every effort to reward you for helping to find them by assuring you at least get some credit. Ordinarily tasks like this would be marked as "failed" or "client error" and given zero credit. But R@h accepts the results, gives some amount of credit for the reported work, and then the project team goes about learning what particular combination of things happens with that specific model that cause it to fail.

---

Rosetta Moderator: Mod.Sense