On a (very) few occasions I end up having to abort a task that has run for over 24 hours and has like 3 months left but the due date is less than a week.

When looking in the event viewer I once saw a "radar pre-leak" warning for one of the period search apps. Possibly these are related. Only 5 bad tasks out of l000 is not a problem. What is a problem is one that can get stuck a long time as they are hard to spot.

This is getting worse. I had 3 tasks in a row (win10x64) that showed %3 GPU usage on gpuz.

I stopped and restarted boinc to see if a reset of the GPUs helped. All that did was set the expected time to complete from 3 months down to a few minutes. The GPU, 1070Ti continued to run at %3 with completion time increasing.

I then aborted all tasks that used only %3 gpu resources. The 4th tasks ran OK. It is using %99 like it should.

I checked my wingmen on the 5 tasks that had errors and they reported did not see a problem like mine. However, they were running linux cuda or the cpu app. No wingmen were running win10 cuda.

I have 9 more tasks queued up. If I continue to have problems will switch back to einstein: 500+ works units with no error.