Red Hat Bugzilla – Bug 65063
(VM)Kernel VM/swap so bad it is nearly unusable
Last modified: 2015-01-04 17:01:52 EST
From Bugzilla Helper:
User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.0rc2) Gecko/20020510
Description of problem:
It functions as far as I can tell, but as long as swap space is turned on, it
seems to swap rather than reclaim pages from the file cache. This very quickly
causes the machine to become unusable when large files are copied or written.
The program which caused the problem was not always killable due to the
interactive X windows being badly swapped out.
More precisely, I suppose this is a problem with the kernel having a policy of
treating file write data as higher priority (or even the same priority) as
As far as I'm concerned, swap is broken in this kernel version.
Version-Release number of selected component (if applicable):
Steps to Reproduce:
1. Copy/FTP a huge file or set of files onto the machine which are larger than
the system's RAM.
2. Try to run some interactive programs.
Actual Results: The machine quickly throws out all it's working sets and gets
Sometimes it was so bad, the interactive shells and/or X windows were swapped
out and would *NOT* come back. This is bad.
Expected Results: It should have been writing out the file data and not
throwing out the working sets of programs.
This sounds quite similar to what was already reported in bug # 60079 and bug #
Funny thing is, the former bug report suggests the problem was resolved in the
rawhide kernel update,
but the kernel update issued for 7.3 should be newer than that bug report.
Hmm. Further testing shows I was smoking something.
It appears what I really saw was the I/O layer paging things in very slowly
under high write load.
Actually, I should be more specific. It pages things in slowly under ANY write
it might be worthwhile to check if "elvtune" can improve this.
elvtune can be used to reduce the read and write latency parameters;
basically larger values for those mean more aggressive reordering of
io requests, which can result in longer latencies and starvation.
Closing (end of lifed).
Latest errata kernels have various fixes in this area btw