You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The PDBReader opens and closes the trajectory file for reading every frame, see. Reopening the files all the time means we are doing direct I/O. This is an issue when MDAnalysis is run on an HPC cluster with 10 to 100 of different processes at the same time.
HPC cluster usually have some kind of heavily buffered distributed filesystem, like the GPFS. These type of filesystems hate direct I/O, they prefer buffered I/O. This would mean keeping a file open for longer period of time and reading from it. They can buffer that. Closing the file breaks their buffering mechanism. The end result is super slow file access for everybody on a cluster.
The solution is to keep the file pointer open. I did the same for the XTC/TRR readers.
The text was updated successfully, but these errors were encountered:
The PDBReader opens and closes the trajectory file for reading every frame, see. Reopening the files all the time means we are doing direct I/O. This is an issue when MDAnalysis is run on an HPC cluster with 10 to 100 of different processes at the same time.
HPC cluster usually have some kind of heavily buffered distributed filesystem, like the GPFS. These type of filesystems hate direct I/O, they prefer buffered I/O. This would mean keeping a file open for longer period of time and reading from it. They can buffer that. Closing the file breaks their buffering mechanism. The end result is super slow file access for everybody on a cluster.
The solution is to keep the file pointer open. I did the same for the XTC/TRR readers.
The text was updated successfully, but these errors were encountered: