Processing performance can significantly decrease when the file cache buffers of a Linux machine are full. While this is true in serial, it is more often a concern when solving large cases in parallel, particularly when using AMD processors. If you see a performance decrease even though the case / machine setup has not changed, that is an indication that the file cache buffers may be to blame.
The filling of the file cache buffers can happen over a period of time as a result of input-output activity. Even after the ANSYS Fluent session is exited, by default the operating system does not free up the file cache buffers immediately (unless the operating system is unable to satisfy a malloc subroutine request with available free memory). During memory allocation for a parallel case, this can result in the allocation of memory from a different NUMA domain, and consequently can have a significant impact on performance.
To resolve this issue on Linux machines, you must first ensure that all of the relevant machines are idle (so that you do not adversely affect any jobs that are running). Then you can clear the file cache buffers by performing one of the following actions:
Include the -cflush option when launching ANSYS Fluent from the
command line.
This option ensures that the file cache buffers are flushed in a separate operation. While this process may take a few minutes to complete (depending on the total memory of the system), it does not require you to have root privileges.
or
Enter the (drop-cache) Scheme command (either in the
ANSYS Fluent console or through your journal file) after launching Fluent but before you
read the case file.
This command will instantaneously clear the pagecaches, dentries, and inodes.
Important: Note that in order to use the (drop-cache) command, you
must have sudo administrative privileges for the /sbin/sysctl
vm.drop_caches=3 command.