Check points are part of the science application, not the client. AutoDock has relatively few check points. It checkpoints when the green line in the Application View reaches the edge of the screen and starts redrawing. HPF check points whenever the Progress indicator increments several tenths of a per cent.

Hi andzgrid,
The problem is that the application program is normally in a state with tens of megabytes of floating point arrays in use. Every once in a while it converges to a solution, and then a small checkpoint can be saved. Then it starts another round of computations. Unlike Rosetta, AutoDock just does not seem to have many natural checkpoints. Rick Alther looked into this in November and gave us an updated version of AutoDock in January, but he was not able to solve the dearth of checkpoints.

lawrencehardin...


I could not have said it better myself...

... just kidding.


Seriously now…
I don’t know but is there something we can learn from looking at how Windows does hibernation? My gut feel tells me that there may be something we can start off from there, some variation thereof. And how does Rosetta get to have (comparatively many) ‘natural’ checkpoints than Autodock? Dearth of checkpoints… Hmmm… This is a wild shot, but here goes: Artificial checkpoints?

The problem with checkpointing isn't identifying and writing all the data to disk. The hard part about checkpointing is the restoration process. Figuring out how to read in the data and set up the data structures at the proper points so that the typically stochastic algorithms resume precisely where they left off. These are very complex calculations with millions of data points where data in one part of the calculation affects other parts later on and so forth.

We choose a location (or locations) in the program where the state is easily known, and more importantly, is easy to resume should we need to do so.

This is on our plate to look into, but I have a five course meal of tasks sitting in front of me

To answer your question about using something similar to Windows hibernation, yes, there are grid systems which use a technique similar to this. There are grid systems where the client application runs inside a virtual machine (think VM Ware). In that case, there is no need for checkpointing since the grid agent can simply take a snapshot of the virtual machine state and resume right from there. The downside to this is it requires virtual machine software on every client and running a virtual machine adds to the RAM/VM overhead. While this is currently impractical for a volunteer grid of devices, it may not be so in the future.