Core21 v0.0.14 has been released!
Posted: Mon Dec 07, 2015 6:41 pm
We rolled out Core21 v0.0.14 last week.
Main features are:
* Bad State errors fixed: Improvements that drastically reduce the high rate of Bad State errors we were seeing with earlier versions of the core, especially with NVIDIA cards.
These Bad State errors were ironically caused by a couple of bugs in the code that checks the simulation for integrity every time a checkpoint file is written. The simulation integrity was solid, but there was a high false positive rate for errors. This should be greatly reduced with 0.0.14.
For NVIDIA cards, this unfortunately comes with a 10% performance regression when PME is in use, but we are compensating by rebenchmarking affected core 21 projects. We hope to release an update that will undo this performance regression soon.
Note that you will still get Bad State errors if your hardware is unstable or highly overclocked.
* More debugging info: When Bad State errors *do* occur, we now bring back more information to help us diagnose these issues.
* Early NaN detection eliminates slowdowns: We previously had some reports of WU slowdowns that were traced to NaNs appearing in the simulation and slowing down the integration loop drastically. Previously, these were not detected until a checkpoint was written. Now, checks are performed much more frequently, hopefully eliminating these slowdowns.
* Minor improvements: We have also made a number of other minor improvements that give us a bit more flexibility and control over what simulations can be run.
Thanks to everyone who helped make this possible, especially Peter Eastman (OpenMM developer extraordinaire!) and all of the testers who helped kick the tires, including (but by no means limited to):
(Updated: I had forgotten in the original post!)
Main features are:
* Bad State errors fixed: Improvements that drastically reduce the high rate of Bad State errors we were seeing with earlier versions of the core, especially with NVIDIA cards.
These Bad State errors were ironically caused by a couple of bugs in the code that checks the simulation for integrity every time a checkpoint file is written. The simulation integrity was solid, but there was a high false positive rate for errors. This should be greatly reduced with 0.0.14.
For NVIDIA cards, this unfortunately comes with a 10% performance regression when PME is in use, but we are compensating by rebenchmarking affected core 21 projects. We hope to release an update that will undo this performance regression soon.
Note that you will still get Bad State errors if your hardware is unstable or highly overclocked.
* More debugging info: When Bad State errors *do* occur, we now bring back more information to help us diagnose these issues.
* Early NaN detection eliminates slowdowns: We previously had some reports of WU slowdowns that were traced to NaNs appearing in the simulation and slowing down the integration loop drastically. Previously, these were not detected until a checkpoint was written. Now, checks are performed much more frequently, hopefully eliminating these slowdowns.
* Minor improvements: We have also made a number of other minor improvements that give us a bit more flexibility and control over what simulations can be run.
Thanks to everyone who helped make this possible, especially Peter Eastman (OpenMM developer extraordinaire!) and all of the testers who helped kick the tires, including (but by no means limited to):
Code: Select all
christianvirtual
davidcoton
bollix47
grandpa
bruce
jimerickson
frodothehobbit
toTOW
Code: Select all
toTOW