...
During run time only few files are involved in NAMD's I/O activities. As long as standard MD runs are carried out, this is unlikely to impose stress on the Lustre file system ($WORK) . For the case of replica-exchange runs, however, the situation can be different. Here we already observed cases where a high frequency of as long as one condition is met. Namely, file metadata operations (file stat, create, open, close, rename) occurred. This has overloaded our Lustre metadata servers resulting in a severe slowdown of the entire Lustre file system. Candidates for heavy metadata file I/O are should not occur at too short time intervals. First and foremost, this applies to the management of NAMD restart files. Instead of having a new set of restart files created several times per second, the NAMD input parameter restartfreq should be chosen such that they are written only every 5 minutes or in even longer intervals. For the case of NAMD replica-exchange runs the situation can be more severe. Here we already observed jobs where heavy metadata file I/O on the individual "colvars.state" files located in every replica's subdirectory . Problems arise when these files get renamed, created, opened, and closed in too short intervalshas overloaded our Lustre metadata servers resulting in a severe slowdown of the entire Lustre file system. Users are advised to set corresponding NAMD input parameters such that each replica performs metadata I/O on these files in intervals longer than at least one secondno shorter than really needed, or, when where affordable, that these files are written only at the end of the run.
...