Suggestions for Poor ext4 performance questions

Kevin Cummings cummings at
Sat Nov 15 06:52:56 UTC 2014

Please don't reply personally.  Please reply to the list.  You posted
your original question to the list, and for the sake of a possible few
others who are following this thread, it is good to keep all replies on
the list.

On 11/14/2014 11:48 PM, George R Goffe wrote:
> Kevin,
> Thank you for your response.
> I have two 4TB file systems and one 6TB filesystem, all ext4 with Journaling enabled. I suspect that as the filesystems becomes inactive, their pages (buffer cache?) in memory are trundled back to disk and then the pages marked as unreferenced. A new access causes this data to be re-read into memory again with a subsequent re-purchase of free memory, hence the delay with the new access. I can not prove this though. I'm sure I'm not stating this correctly but, my hypothesis is that since this is a demand paged system, other activity eventually causes these unreferenced pages to be stolen. System memory is at 8G now. I've been considering upgrading to 16G but the expense is something I'd like to avoid if possible. I was looking for a kernel parameter to tweak which might affect this behavior. I could be totally wrong of course.

This is information that should have been sent in your original request
for help.  You don't want to tune the file-system, you want to tune the
system's use of the paging of the file system.  It helps to be clear so
you don't go running down unnecessary tangents.

What I can suggest is that you look into a kernel parameter known as
"swappiness".  It can help determine when pages are paged back to disk.
 In your case, you want to keep them in memory (disk cache) for as long
as possible before they get written out to disk.  swappiness can help
control that.  For more information google "linux swappiness".  It even
has its own Wikipedia page....

> Regards and THANKS for your help/time,
> George...

Kevin J. Cummings
kjchome at
cummings at
cummings at
Registered Linux User #1232 (

More information about the users mailing list