Some users have been reporting that some workunits are taking significantly longer to run than what wingmen are reporting -- which is a pretty weird bug. See this forum thread.

For a status update, what I believe is happening is that certain parameters to the backpropagation is causing this issue on some operating systems/architectures. I've whipped up a page to track these "monster" workunits, which can be found here. This page shows all workunits with results that had runtimes more than 4x apart from each other. Quite a few of these had initial learning rates set at 1e-08 which is extremely low. I'm thinking this might be part of the problem so I've updated things server side to prevent the initial learning rate from being any lower that 1e-05.

I'm hoping that with this change, newly generated workunits should not have the problem anymore. If you're still seeing it please let me know. If you see it happen with a workunit that doesn't have a very low initial learning rate, please let me know as well as I'll need to do some more digging.

More...