From Fedora Project Wiki

< FWN‎ | Beats

Revision as of 04:02, 2 November 2009 by Dale (talk | contribs) (→‎KSM Tuning)


In this section, we cover discussion of Fedora virtualization technologies on the @fedora-virt and @virt-tools lists..

Contributing Writer: Dale Bewley

Fedora Virtualization List

This section contains the discussion happening on the fedora-virt list.

KSM Tuning in Fedora 12

Mark McLoughlin described[1] the default state of KSM[2] on Fedora systems. "For Fedora 13, it'll be off by default in the kernel and the recommended way of switching it on is with 'chkconfig ksm on'" "For Fedora 12, it's on by default in the kernel, 'chkconfig ksm on' just changes max pages and the only way of disabling it is by manually writing zero to /sys/kernel/mm/ksm/run."

At release of Fedora 12 the kernel will default to a maximum of 2000 merged memory pages. A future F12 kernel update to 2.6.32 will likely disable KSM by default. To take advantage of KSM in Fedora 12, the ksm service must be enabled:

sudo chkconfig ksm on

Mark McLoughlin also noted[3] The maximum number of pages which may be merged defaults to half of the system memory, and may also be manually defined in /etc/sysconfig/ksm. "Here's the logic we have in the init script[4]:"

  # unless KSM_MAX_KERNEL_PAGES is set, let ksm munch up to half of total memory.
  default_max_kernel_pages () {
      local total pagesize
      total=`awk '/^MemTotal:/ {print $2}' /proc/meminfo`
      pagesize=`getconf PAGESIZE`
      echo $[total * 1024 / pagesize / 2]

Justin Forbes points out[5] "The limit to half of total memory is because ksm pages are unswappable at this time. To be fixed in a future kernel."

A second service, ksmtuned, may also be enabled. Ksmtuned regulates how aggressively the system will attempt to merge pages. Parameters such as how many pages to scan before sleeping and how long to sleep may be configured in /etc/ksmtuned.conf.

Memory pages must be flagged as mergable before KSM will scan them looking for duplicates. At present only Qemu pages will be marked as such. As described in the kernel docs[6], the effect of KSM system memory may be examined in /sys/kernel/mm/ksm. "A high ratio of pages_sharing to pages_shared indicates good sharing, but a high ratio of pages_unshared to pages_sharing indicates wasted effort."