The new core has a number of improvements that will help us better support COVID-19 drug discovery collaborations, especially the COVID Moonshot (http://postera.ai/covid).
New Features:
- Now built on the most up-to-date release of OpenMM: OpenMM 7.4.2
- Average performance (in ns/day) is now printed out to the client log at the end of the WU.
- JSON viewer frames are now written every 1% by default, and project managers can now flexibly control the viewer frame interval for each project.
- Checkpoints are now written every 5% by default, and project managers can now flexibly control the checkpoint interval independent of the XTC frame write interval for each project.
- Checkpoint, XTC, and viewer frame intervals are now written to client log
- Adds experimental support for Intel GPUs
- Slight performance improvement on some GPUs by enabling separate PME streams by default
- Issues causing some AMD cards (AMD RX 460/470/480/570/580/590) to fail with
was fixed in OpenMM (openmm/openmm#2608) and included in OpenMM 7.4.2
Code: Select all
Error invoking kernel sortShortList: clEnqueueNDRangeKernel (-5) - Project managers now have control over force and energy RMSE error checking thresholds, which should allow project managers to reduce RMSE energy and force errors for large projects by setting these thresholds larger.
- Fixes bugs in resuming from projects that use CustomIntegrators (like 114xx for the COVID Moonshot) and correctly rewinds globals.csv files to ensure integrity of data when core is paused/resumed
Some short projects may note slightly reduced performance in exchange for having to backtrack less when NaNs are encountered due to more frequent checkpointing. We hope to tune this and further improve behavior in future, and have some ideas as to how we can keep the GPU running while the CPU is performing sanity checks and checkpoints in future core22 updates.
We're also hoping to make further performance updates in the coming weeks. Stay tuned!
~ John Chodera and the core22 team
