On 11/10/06, Alan Ackerman <[EMAIL PROTECTED]> wrote:

Mostly VM uses LRU algorithms -- the VDISK page is NEW to VM when it is f
irst written, so it will be retained longer than unchanged pages in the Linux 
virtual machine. As
others have pointed out, the actual algorithms are more complicated than LRU -- 
and they tend
 to favor keeping VDISK pages in memory even longer.

That's the problem with "opinions" based on belief rather than
measurements. For most installations VDISK pages are not favored over
primary address space pages. Not even on older releases of z/VM.

What I remember reading, but have not tried, is the suggestion that you h
ave multiple levels of swap space, a small VDISK and larger minidisks to
back it up. That might help in this case.

The need for multiple layers of swap space is to help z/VM understand
what to keep resident. A small VDISK will be referenced more often and
thus remains in storage. Sizing of the virtual machine and the
different VDISKs needs measurements. Because VDISK is fast and cheap,
the performance penalty for a changed application is not too much.

The total amount of swap space is an application issue. Linux wants to
see the swap space to feel happy to allow allocation of virtual
memory, but much of it will not be used by the application. So you
define swap space that will probably never be referenced.

I do not see a reason to put that "just in case" swap space on real
disk. If it is not referenced it makes no difference, and if it is
referenced (because of bad sizing or change of application) the real
I/O will make it painful for everyone. Instead, give all that disk
space to z/VM as paging devices and let it be used to support the
paging subsystem.

Clearly people have been burned by skipping that last step: give CP
enough paging space to support virtual machines and VDISK. If you're
not able to allocate z/VM paging large enough to hold all entire
virtual machines and VDISKs (but go and look what you have allocated
for swap on real disk), you need to monitor the usage carefully.
Checking page space utilization should be on your performance
checklist (i.e. monitor alerts). Since you need to be under 50% anyway
(for efficient block paging) there is a lot of time to see the trend
change.

Rob
--
Rob van der Heij
Velocity Software, Inc
http://velocitysoftware.com/

Reply via email to